Sources Directory

Type: kb/types/index.md

← Parent

"Creative Thinking" (snapshot)
/problem-first: a simple skill to invert bad ideas (snapshot)
A harness for every task: dynamic workflows in Claude Code (snapshot)
A Mini Exercise on the Mismanaged Geniuses Hypothesis (RLMs on LongCoT) (snapshot) - Alex Zhang and Omar Khattab's LongCoT-mini RLM prompt-tuning case study, where guardrails and graph-aware decomposition raise GPT-5.2 RLM performance from 50.6% to 65.6%
A Model for Types and Levels of Human Interaction with Automation (snapshot) - Full-text capture of the Parasuraman-Sheridan-Wickens framework for allocating automation across four function types and graded levels of human interaction.
A new way to think about composing skills to increase leverage: Skill Graphs 2.0 (snapshot)
A-Mem: Agentic Memory for LLM Agents (snapshot)
A-MEM: Learning Operations Analysis (note) - Dissects A-MEM's four fully-automatic operations (construct, link, evolve, retrieve) — all accretive, none curative — identifying the missing vocabulary (delete, split, reorganize, assess quality) that separates accumulation from curation
ACM: Agentic Context Management for Long Horizon Tasks (snapshot) - Agent-controlled, lossless context compression backed by external memory, with post-training results on long-horizon search and coding tasks
Adaptation of Agentic AI: A Survey of Post-Training, Memory, and Skills (snapshot) - Survey taxonomy of agentic AI adaptation across post-training, memory, skills, and agent/tool optimization paradigms
Agent Behavioral Contracts for Reliable Agents (snapshot)
Agent Harness for Large Language Model Agents: A Survey (snapshot) - Survey framing the LLM agent harness as a six-component runtime governance layer and primary determinant of deployed agent reliability
Agent Workflow Memory (snapshot) - AWM paper on inducing reusable web-agent workflows online or offline, with WebArena and Mind2Web evaluations against exemplar prompting and workflow baselines.
Agentic Code Reasoning (snapshot) - Semi-formal reasoning templates (explicit premises, execution traces, formal conclusions) improve LLM code verification by 5-12pp across patch equivalence, fault localization, and code QA tasks
Agentic Memory for LLM Agents (snapshot)
Agentic Note-Taking 23: Notes Without Reasons (snapshot)
Agents Explore but Agents Ignore: LLMs Lack Environmental Curiosity (snapshot) - Solution-injection benchmark paper showing agents often discover explicit task solutions but fail to exploit them.
AI Agents: Four Skills More Important Than a Good Prompt (snapshot) - English translation of Mezha article applying Anthropic's AI Fluency framework to agent use: delegation, description, discernment, and diligence as operator skills beyond prompting.
AI Components for a Deterministic System (An Example) (snapshot)
An Enigma of Artificial Reason (snapshot) - Paper introducing VAIR, a benchmark showing large reasoning models can produce correct answers while failing to evaluate invalid reasoning that reaches those answers
Andrej Karpathy talks about "Claws" (snapshot)
Artifacts as Memory Beyond the Agent Boundary (snapshot) - RL paper formalizing environmental artifacts as externalized memory that can reduce agent history capacity.
Ashby, Design for a Brain — ultrastability (source-review) - Ashby's ultrastable system — the cybernetic precedent for a second loop that reorganizes a system when essential variables leave viable limits, retaining by equilibrium rather than by an evaluator
Autogenesis: A Self-Evolving Agent Protocol (snapshot) - Autogenesis protocol specification and benchmark evidence for versioned, auditable evolution of prompts, agents, tools, environments, and memory
Automated linking improves retrieval but may degrade navigability (note) - Triangulates A-MEM, Notes Without Reasons, and the open-problem note — automated linking improves retrieval (QA benchmarks) but degrades navigability (agent trust in link infrastructure); the distinction is adjacency versus connection
Autonomy and Autopoiesis (snapshot) - Varela's distinction of autonomy and autopoiesis from looser notions of self-organization
Autoreason: Self-Refinement That Knows When to Stop (snapshot) - Self-refinement paper that makes "do nothing" a first-class candidate via blind fresh-agent Borda tournaments, finding gains mostly in mid-tier models with a generation-evaluation gap
Beyond Transformers: Sudoku Bench (snapshot) - Pathway's BDH model achieves 97.4% accuracy on extreme Sudoku while leading LLMs score 0%, using the gap as evidence that transformer architecture has fundamental limits for constraint-satisfaction reasoning and arguing for post-transformer latent-space models.
Beyond “Not Novel Enough”: Enriching Scholarly Critique with LLM-Assisted Feedback (snapshot) - LLM-assisted scholarly novelty-assessment paper using related-work retrieval and structured comparison against human novelty reviews.
Build Systems à la Carte (snapshot) - Framework unifying Make, Excel, Shake, Bazel, Buck, and Nix as scheduler+rebuilder combinations, formalizing minimality, dynamic dependencies, and early cutoff.
Building a Good Vertical Agent (snapshot)
Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning (snapshot) - Agentic automata-learning benchmark evaluating whether tool-calling LLM agents infer hidden DFAs through membership and equivalence queries.
Can LLMs Perform Deep Technical Comprehension of Computer Architecture Papers? (snapshot) - Gauntlet study comparing independent expert-persona review plus disagreement-preserving synthesis against human and single-agent analysis of computer architecture papers
Causal inference using invariant prediction: identification and confidence intervals (snapshot) - Peters, Buehlmann, and Meinshausen use invariance across environments and interventions to identify causal predictors and confidence intervals
Causal-learn: Causal Discovery in Python (snapshot) - Zheng et al. describe causal-learn, a Python library collecting causal discovery methods for inferring causal structure from observational data under method-specific assumptions
Claude Fable 5 Made Most of My Agent Scaffolding Obsolete. Here's What Survived. (snapshot)
Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents (snapshot) - Co-Harness alternates validated harness optimization with model fine-tuning that distills trajectories produced by the improved scaffolding.
Coding Agents are Effective Long-Context Processors (snapshot) - Benchmark paper arguing coding agents process long contexts better by turning text corpora into file-system-native tool workflows rather than latent attention or fixed retrieval
Cognee: Knowledge Engine for AI Agent Memory (snapshot)
Comparative Research: Emotional Prompting, Stress & Sycophancy (snapshot) - RESEARCH.md snapshot for Gentle-Coding: an annotated bibliography of emotional-prompting, stress, and sycophancy research organized by sub-network, inference, and multi-agent operational layers.
Components of A Coding Agent (snapshot) - Raschka's breakdown of the six architectural components of a coding agent harness — distinguishing the harness from the model and arguing that context quality drives apparent model quality.
Computational Reflection (snapshot) - Pattie Maes's general definition of computational reflection through causally connected self-representation, with procedural, declarative, theory-relative, implicit, and explicit variants
Concept Bottleneck Models (snapshot) - Models that predict human-understandable concepts as an intermediate bottleneck before the final task prediction, so a human can intervene on a wrong concept at test time and correct downstream errors without retraining
Concepts and Experiments in Computational Reflection (snapshot) - Pattie Maes's concise early definition of computational reflection and 3-KRS demonstration of explicit, uniform object-oriented meta-objects
Context Engineering for AI Agents in Open-Source Software (snapshot)
Context Is What You Need: The Maximum Effective Context Window for Real World Limits of LLMs (snapshot) - Empirical study defining and measuring Maximum Effective Context Window (MECW) across 11 frontier LLMs — finds MECW is drastically smaller than advertised MCW, shifts by task type, and that large context windows cause hallucination rates to approach 100%.
Context providers: the missing layer between agents and tools (snapshot)
Continual Harness: Online Adaptation for Self-Improving Foundation Agents (snapshot) - Reset-free embodied-agent study in which an LLM refiner edits prompts, sub-agents, skills, and memory online, with capability-dependent Pokémon results and joint model-harness training.
Continual Learning in Token Space (snapshot) - Letta reframes continual learning for agents as optimization over learned context rather than weights, arguing token-space memory is the primary transferable substrate for long-lived agents
ConvexBench: Can LLMs Recognize Convex Functions? (snapshot)
Cybernetics of Cybernetics (snapshot) - Von Foerster's primary second-order move from observed systems to observing systems and the observer's purposes
Dario Amodei — "We are near the end of the exponential" (snapshot) - Anthropic CEO's capability-timeline predictions — verifiable domains get confident timelines, unverifiable ones get hedged, implicitly confirming oracle-strength thesis
Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents (snapshot) - Darwin Gödel Machine evolves a lineage archive of self-editing coding agents under benchmark selection while keeping the exploration controller fixed
Defining Self-adaptive Systems: A Systematic Literature Review (snapshot) - Full-text capture of a systematic review of formal definitions and characterization requirements for self-adaptive systems.
DiscoverPhysics: Benchmarking LLMs for Out-of-the-Box Scientific Thinking (snapshot) - Benchmark testing whether LLM agents can discover novel physics laws in simulated worlds with exotic physics through iterative experimentation, rather than recalling known physics — best agents solve only about half the worlds and struggle with latent structure discovery.
Do Agent Optimizers Compound? A Continual-Learning Evaluation on Terminal-Bench 2.0 (snapshot) - Two-phase Terminal-Bench evaluation separates static optimizer gains from transfer and re-optimization, attributing RELAI-VCL's advantage to in-loop regression control.
Does Liking Yellow Imply Driving a School Bus? Semantic Leakage in Language Models (snapshot) - Defines and measures semantic leakage — undue prompt-to-generation association from learned concept links — across 13 GPT and Llama models via control/test prompt pairs and Leak-Rate metric; instruction-tuned models leak more.
DoWhy: Addressing Challenges in Expressing and Validating Causal Assumptions (snapshot) - Sharma et al. argue causal estimates depend on explicit assumptions and describe DoWhy support for expressing and partially validating those assumptions
DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning (snapshot) - DreamCoder paper — wake-sleep Bayesian program synthesis that grows a domain-specific language of reusable abstractions alongside neural search guidance, producing compositional, interpretable, transferable symbolic knowledge from experience.
Effective harnesses for long-running agents (snapshot) - Anthropic's long-running-agent experiments show why compaction needs persistent progress artifacts, incremental work, structured requirements, and end-to-end verification.
Emergent Analogical Reasoning in Transformers (snapshot) - Transformer analogy paper that treats analogical reasoning as relational-structure alignment plus functor-like vector transformation
Emerging Markdown Formats That Shape Coding Agent Behavior (snapshot) - Map distinguishing standing rules, repeatable procedures, versioned plans, domain files, and machine-local memory in agent-ready repositories
EsoLang-Bench (snapshot) - OOD code benchmark using esoteric languages to separate transferable reasoning from benchmark memorization and contamination
Ethics and Second-Order Cybernetics (snapshot) - Von Foerster's observer-inclusive account of responsibility, undecidable questions, and second-order cybernetics
Evaluating Long-Context Reasoning in LLM-Based WebAgents (snapshot) - Benchmark showing LLM-based web agents fail badly under long context with injected irrelevant task sequences — success rates drop from 40-50% to under 10% at 150k tokens, with loop and lost-objective failures dominating; implicit RAG provides only modest relief.
Every Good Regulator of a System Must Be a Model of That System (snapshot) - Conant and Ashby's theorem that optimal simple regulation requires a model of the regulated system
Everything you need to know about LLM memory (snapshot) - Notion essay arguing that LLM memory needs retrieval, salience, summarization, forgetting, and memory objects rather than raw chat logs.
Evolving self-reference: matter, symbols, and semantic closure (snapshot) - Pattee's semantic-closure account of matter-symbol complementarity in living systems capable of open-ended evolution
Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering (snapshot) - Survey paper framing LLM agent progress as externalization into memory, skills, protocols, and harness engineering rather than only stronger model weights.
FALSIFYBENCH: Evaluating Inductive Reasoning in LLMs with Rule Discovery Games (snapshot) - Wason-2-4-6-style benchmark finding that LLM success at rule discovery is driven primarily by negative testing (falsification-seeking), not confirmation-seeking
From Agent Loops to Structured Graphs: A Scheduler-Theoretic Framework for LLM Agent Execution (snapshot) - Position paper recasting the Agent Loop as a single-ready-unit scheduler and proposing Graph Harness (SGH) — a static-DAG execution model with immutable plan versions, three-layer separation, and bounded three-level recovery.
From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence (snapshot)
From Human Memory to AI Memory: A Survey on Memory Mechanisms in the Era of LLMs (snapshot) - Survey proposing a 3D-8Q taxonomy (object/form/time) for LLM memory mechanisms, mapping human memory types (sensory, working, explicit, implicit) to AI implementations across eight quadrants — useful for comparing how different systems position their memory architectures.
Frontis-MA1: Training an AI4AI Model toward Recursive Self-Improvement in MLE (snapshot) - Open full-stack MLE system aligns execution-grounded post-training with four program-evolution operators and long-horizon evolutionary search
Gentle Coding Framework (snapshot) - README snapshot for Gentle-Coding, a community research hub proposing Gentle Mindset, Defined Winning Condition, and Safety-Token as prompt-framing concepts for AI coding and reasoning tasks.
Gentle-Coding (snapshot) - Proof-of-Concept.md snapshot for Gentle-Coding: the original small-scale PoC report comparing authoritarian vs. gentle prompt framing across matrix, sequence, and riddle tasks on six cloud models.
Geometry of Knowledge Allows Extending Diversity Boundaries of Large Language Models (snapshot) - Proposes a plug-in latent-conditioning framework that traverses semantic manifolds to expand LLM generation diversity without fine-tuning, with empirical gains on NoveltyBench and the Alternative Uses Test.
GIANTS: Generative Insight Anticipation from Scientific Literature (snapshot) - GIANTS benchmark and 4B model for predicting downstream scientific insights from two cited parent-paper summaries
Graphiti: Temporal Knowledge Graph for AI Agents (snapshot)
Gödel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements (snapshot) - Schmidhuber's Gödel machine — a self-referential problem solver that rewrites any part of its own code, including its proof searcher, only once an embedded proof searcher proves the rewrite raises axiomatized utility
Harness Engineering for Self-Improvement (snapshot) - Lilian Weng's research synthesis treats harness code, workflow, context, memory, and evaluators as an executable self-improvement surface
Harness Engineering Is Cybernetics (snapshot)
Harness Engineering: Leveraging Codex in an Agent-First World (snapshot)
Harness Handbook: Making Evolving Agent Harnesses Readable, Navigable, and Editable (snapshot) - Behavior-centric, source-linked handbook representation and progressive-disclosure workflow for localizing distributed implementation sites in evolving agent harnesses.
Harness Updating Is Not Harness Benefit (snapshot) - Controlled agent-evolver cross-pairing separates producing persistent harness updates from activating and following them during later task solving.
How I built a self-improving software factory (snapshot)
How Is LLM Reasoning Distracted by Irrelevant Context? An Analysis Using a Controlled Benchmark (snapshot) - Introduces GSM-DC, a controlled benchmark using symbolic DAGs to systematically measure how irrelevant context degrades LLM reasoning — quantifies power-law error scaling with distractor count, and shows Hard-IC training plus PRM-guided tree search are the most effective robustness interventions.
How to Build an AI Second Brain With Claude and Obsidian That Gets Smarter Every Day (Full Guide) (snapshot)
How to build your own agent harness??? (snapshot)
How to Recursively Improve Your Agents (snapshot) - Agno walkthrough deriving probes from an agent specification and usage traces, then editing instructions, tools, and parameters until the probes pass
How we built our knowledge base (snapshot)
Human Bottlenecks (snapshot) - Fernando Borretti's argument that AI's transformative potential is capped by internal human bottlenecks — executive function, intelligence, and foundational knowledge — that software cannot substitute for.
Human Routers of Machine Words (snapshot) - Fernando Borretti's polemic arguing that outsourcing writing to AI is contemptible because writing is thinking — not transcription — and delegating it skips the cognitive work that makes ideas real.
Huxley-Gödel Machine (snapshot) - ICLR 2026 Huxley-Gödel Machine paper proposing clade-metaproductivity as a better search signal for self-improving coding agents than immediate benchmark score
HyperAgents (snapshot) - Self-referential DGM variant makes task and meta-agent code jointly editable, with cross-domain transfer and compounding self-improvement claims
Improving AI Skills with autoresearch & evals-skills (snapshot)
In Search of Lost Domain Generalization (snapshot) - DomainBed testbed paper showing well-tuned ERM matches or beats specialized domain-generalization algorithms once a model selection strategy is made explicit
in-toto: Providing farm-to-table guarantees for bits and bytes (snapshot) - USENIX Security '19 paper introducing in-toto, a framework that cryptographically verifies the integrity of each step in a software supply chain from source to deployment via a signed layout and per-step link metadata.
Infinite midwit (snapshot) - Adam Mastroianni's objective-vs-subjective intelligence framing for why AI competence and benchmarks still miss taste, wisdom, and idea selection.
Ingest Report: SkillOpt: Executive Strategy for Self-Evolving Agent Skills (ingest-report) - SkillOpt paper showing validation-gated text-space optimization of compact agent skills as readable deploy-time learning around frozen models
Ingest: "Creative Thinking" (ingest-report) - Shannon's 1952 lecture cataloguing six explicit problem-solving operators (simplification, analogy, restatement, generalization, structural analysis, inversion) as a portable creative toolkit
Ingest: /problem-first: a simple skill to invert bad ideas (ingest-report) - Practitioner report on a PM skill that forces solution ideas back into problem statements, useful as a process-structure example for agent skills
Ingest: A harness for every task: dynamic workflows in Claude Code (ingest-report) - Anthropic practitioner account of dynamic workflows in Claude Code: model-authored ephemeral JS orchestrators that spawn and coordinate sub-agents
Ingest: A Mini Exercise on the Mismanaged Geniuses Hypothesis (RLMs on LongCoT) (ingest-report) - LongCoT-mini RLM case study where trace-extracted prompt tips, guardrails, and sub-answer checking improved graph-structured compositional reasoning
Ingest: A Model for Types and Levels of Human Interaction with Automation (ingest-report) - Parasuraman, Sheridan, and Wickens provide a human-centered function-allocation matrix: four automation stages crossed with manual-to-autonomous levels and evaluated by performance consequences.
Ingest: A new way to think about composing skills to increase leverage — Skill Graphs 2.0 (ingest-report) - Sakhuja's practitioner reframe of skill graphs into a three-tier compositional hierarchy (atoms/molecules/compounds) driven by the reliability ceiling of deep skill chains and the 'brain RAM' bottleneck in parallel-agent supervision
Ingest: A Scheduler-Theoretic Framework for LLM Agent Execution (ingest-report) - Position paper formalising LLM agent execution as a scheduler; corroborates the KB's clean-model orchestration cluster and supplies ready-set cardinality as a quantitative axis
Ingest: A-MEM: Agentic Memory for LLM Agents (ingest-report) - Zettelkasten-inspired flat agent memory with embedding linking and LLM-driven evolution — benchmark success without curation operations or inspectable links
Ingest: ACM: Agentic Context Management for Long Horizon Tasks (ingest-report) - ACM improves three benchmarks inside a fixed two-tool decomposition, but neither tests that decomposition's scope nor shows lossless active context
Ingest: Adaptation of Agentic AI (ingest-report) - Survey mapping agentic adaptation across A1/A2 agent training, T1/T2 tool adaptation, memory, skills, and dynamics-aware evaluation
Ingest: Agent Behavioral Contracts for Reliable Agents (ingest-report) - ABC extends Design by Contract to agents with probabilistic compliance, Lyapunov drift bounds, hard/soft constraints, typed recovery, and a YAML contract DSL
Ingest: Agent Harness for Large Language Model Agents (ingest-report) - Survey defining the LLM agent harness as execution loop, tool registry, context manager, state store, lifecycle hooks, and evaluation interface
Ingest: Agent Workflow Memory (ingest-report) - AWM paper showing web agents can learn reusable prompt workflows from successful trajectories, with online induction helping most as train-test domain gaps widen.
Ingest: Agentic Code Reasoning (ingest-report) - Explicit-premise reasoning templates with execution traces and formal conclusions improve code verification by 5–12 points, supporting structure as interpretation control at 2.8× step cost
Ingest: Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for LLM Agents (ingest-report) - RL-trained unified LTM/STM memory policy for LLM agents — confirms memory management is learnable when task-completion oracles exist, but operates on opaque weights and low-reach facts
Ingest: Agentic Note-Taking 23: Notes Without Reasons (ingest-report) - First-person agent testimony that propositional link semantics differ in kind from embedding adjacency, with a Goodhart corruption argument and an unresolved curation-scaling question
Ingest: Agents Explore but Agents Ignore (ingest-report) - Solution-injection paper showing agents observe explicit task solutions in context but often fail to integrate them into action.
Ingest: AI Agents: Four Skills More Important Than a Good Prompt (ingest-report) - AI Fluency explainer mapping operator competence beyond prompts onto delegation, description, discernment, and diligence for agent work.
Ingest: AI Components for a Deterministic System (An Example) (ingest-report) - Evans argues that separating modeling (schema creation) from classification (schema application) tames LLM non-determinism — a practitioner case study of constraining via taxonomy freezing
Ingest: An Enigma of Artificial Reason (ingest-report) - VAIR paper showing large reasoning models can solve math problems while failing to evaluate invalid reasoning that reaches correct answers
Ingest: Andrej Karpathy talks about "Claws" (ingest-report) - Willison and Karpathy framing "Claw" as a term of art for local persistent AI-agent systems with scheduling, context, tools, and personal-hardware execution.
Ingest: Artifacts as Memory Beyond the Agent Boundary (ingest-report) - RL paper formalizing environment-side artifacts as externalized memory and testing memory by capacity/performance counterfactuals.
Ingest: Autogenesis: A Self-Evolving Agent Protocol (ingest-report) - Autogenesis makes the editable boundary of agent self-improvement explicit, but its benchmarks validate selected prompt, solution, and agent edits rather than the full protocol or safety claims
Ingest: Autonomy and Autopoiesis (ingest-report) - Terminological guardrail separating autopoiesis, organizational closure, and general autonomy
Ingest: Autoreason: Self-Refinement That Knows When to Stop (ingest-report) - Autoreason paper showing self-refinement improves only when candidate synthesis is paired with blind comparative judging and incumbent survival, with gains concentrated in the generation-evaluation gap
Ingest: Beyond "Not Novel Enough" (ingest-report) - Assesses an LLM-assisted scholarly novelty-review paper as evidence for soft-oracle hardening, human-analysis-first evaluator design, and separating reasoning alignment from conclusion agreement
Ingest: Beyond Transformers: Sudoku Bench (ingest-report) - Company Sudoku benchmark reports 97.4% for an undisclosed BDH model versus 0% for LLMs; weak methodology but a third domain suggesting architectural limits in constraint satisfaction
Ingest: Build Systems à la Carte (ingest-report) - How the scheduler×rebuilder build-systems framework grounds the KB's derived-artifact freshness machinery (staleness, verifying traces, recompute-vs-store)
Ingest: Building a Good Vertical Agent (ingest-report) - BrainsAndTennis on vertical-agent quality as task-distribution-aware context compression, with L1/L2/L3 cache tiers for prompts, curated specs, and raw references
Ingest: Can LLM Agents Infer World Models? (ingest-report) - Agentic automata-learning paper showing that hard-oracle interaction tasks can separate LLM-agent evidence collection, hypothesis construction, and final world-model success.
Ingest: Can LLMs Perform Deep Technical Comprehension of Computer Architecture Papers? (ingest-report) - Gauntlet paper shows independent expert perspectives plus disagreement-preserving synthesis improve technical-paper critique, while calibration and confident-error rejection remain unsolved
Ingest: Causal inference using invariant prediction (ingest-report) - Invariant prediction grounds reach assessment by treating cross-environment invariance as evidence for causal predictors rather than fitted correlations
Ingest: Causal-learn: Causal Discovery in Python (ingest-report) - Causal-learn grounds the observational-causal-discovery route to reach assessment, with the important limitation that discovery is assumption-relative
Ingest: Circling Back — Clearing Up Myths About the Deming Cycle (ingest-report) - Moen & Norman's history of the PDSA cycle as scientific method in industry — an applied-lineage witness for the discovery lifecycle and improvement-loop notes.
Ingest: Claude Fable 5 Made Most of My Agent Scaffolding Obsolete (ingest-report) - claude-workstream-kit announcement arguing that stronger models relax model-management scaffolding but make project-scoped, git-versioned active work state more important
Ingest: Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents (ingest-report) - Co-Harness provides a provisional three-form learning case: validated harness repair alternates with fine-tuning on trajectories produced by the repaired harness
Ingest: Coding Agents are Effective Long-Context Processors (ingest-report) - Benchmark paper claiming coding agents beat RAG and context scaling on long-context tasks by using filesystem-native search, slicing, and scripting
Ingest: Cognee: Knowledge Engine for AI Agent Memory (ingest-report) - Pipeline-first knowledge engine with custom Pydantic schemas for LLM entity extraction, poly-store graph+vector design, and an undersized enrichment phase that concretely marks the boundary between automatable extraction and open enrichment problems
Ingest: Components of A Coding Agent (ingest-report) - Practitioner decomposition of coding agent harnesses into six named components, with the central claim that apparent model quality is really context quality — independent convergent evidence for the KB's context-efficiency thesis.
Ingest: Computational Reflection (ingest-report) - Primary vocabulary anchor for computational reflection: causally connected self-representation, its representational variants, and its limits
Ingest: Concept Bottleneck Models (ingest-report) - Architecture-side evidence that a parametric model's intermediates can be made inspectable and correctable by design rather than by interpretability tooling — but only per-inference, not retained
Ingest: Concepts and Experiments in Computational Reflection (ingest-report) - Earlier concise statement of Maes's causal reflection definition, with 3-KRS as an object/meta-object implementation and granularity case
Ingest: Context Engineering for AI Agents in Open-Source Software (ingest-report) - Study of context files across 466 OSS projects identifies five constraint styles, add-then-modify evolution, and 50% stagnation, supplying naturalistic evidence for Commonplace's constraining theory
Ingest: Context providers: the missing layer between agents and tools (ingest-report) - Ashpreet Bedi's ContextProvider pattern: source-scoped sub-agents collapse many raw tools into query/update surfaces to reduce tool-context pollution
Ingest: Continual Harness: Online Adaptation for Self-Improving Foundation Agents (ingest-report) - Continual Harness adds a reset-free all-three-form learning case, bounded by direct edit adoption, low artifact reuse, and a fixed embodied-game decomposition
Ingest: Continual Learning in Token Space (ingest-report) - Letta reframes continual learning as optimizing learned context rather than weights, but the KB's stronger frame is weight space versus repo artifacts, including codified procedures
Ingest: ConvexBench: Can LLMs Recognize Convex Functions? (ingest-report) - Benchmark proving LLM compositional reasoning collapses with depth (not token count), recovered by recursive decomposition with focused context — quantitative evidence for scheduling model predictions
Ingest: Cybernetics of Cybernetics (ingest-report) - Primary second-order-cybernetics vocabulary for observer-inclusive boundaries, purposes, and observing systems
Ingest: Dario Amodei — "We are near the end of the exponential" (ingest-report) - Anthropic CEO's capability-timeline predictions implicitly confirm oracle-strength thesis — verifiable domains (coding, math) get confident timelines while unverifiable domains (novel writing, science) get hedged ones
Ingest: Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents (ingest-report) - DGM declares its own departure from the proof-governed Gödel machine, but its mechanism is looser than a benchmark gate: viability alone admits a child to a monotonic archive, and score only weights reproduction
Ingest: Defining Self-adaptive Systems: A Systematic Literature Review (ingest-report) - Petrovska, Erjiage, and Kugele's systematic review quantifies the definition gap in self-adaptive-systems research and identifies uncertainty and goal semantics as missing formal dimensions.
Ingest: Design Theory in Information Systems (ingest-report) - Gregor's initial Type V account grounds design-and-action theory while separating prescriptive content from Commonplace's proposed operator condition
Ingest: DiscoverPhysics: Benchmarking LLMs for Out-of-the-Box Scientific Thinking (ingest-report) - DiscoverPhysics authors 22 counterfactual physics worlds to defeat recall; agents that predict trajectories well still explain badly — a third target construction and an accuracy/explanation split
Ingest: Do Agent Optimizers Compound? A Continual-Learning Evaluation on Terminal-Bench 2.0 (ingest-report) - Two-phase optimizer study exposes transfer and re-optimization failures, but its compounding claim lacks a fresh-start causal control and rests on a fixed Terminal-Bench decomposition.
Ingest: Does Liking Yellow Imply Driving a School Bus? Semantic Leakage in Language Models (ingest-report) - Semantic leakage — undue prompt-to-generation association from unrelated context — measured by control/test Leak-Rate across 13 models; instruction-tuned models leak more
Ingest: DoWhy: Addressing Challenges in Expressing and Validating Causal Assumptions (ingest-report) - DoWhy grounds the assumption boundary for causal reach assessment: causal estimates require declared assumptions and only partial validation
Ingest: DreamCoder: growing interpretable knowledge with wake-sleep Bayesian program learning (ingest-report) - DreamCoder grows a symbolic library under a description-length gate; lands as the middle case between ungated and proof-gated self-improvement, and the formal-domain limit of mechanical abstraction
Ingest: Effective harnesses for long-running agents (ingest-report) - Anthropic practitioner report separating context continuation from work continuation through persistent progress state, structured requirements, incremental commits, and end-to-end verification
Ingest: Emergent Analogical Reasoning in Transformers (ingest-report) - Transformer analogy paper linking analogical transfer to relational-role alignment, with useful evidence for discovery, reach, and cognitive-analogy transfer methodology
Ingest: Emerging Markdown Formats That Shape Coding Agent Behavior (ingest-report) - Practitioner taxonomy separating standing rules, task procedures, reviewed plans, domain context, and machine-local memory in agent-ready repositories
Ingest: EsoLang-Bench (ingest-report) - Esoteric-language code benchmark arguing standard coding scores mostly measure pretraining fit, with interpreter feedback beating textual critique on OOD tasks
Ingest: Ethics and Second-Order Cybernetics (ingest-report) - Observer participation and responsibility as ethical consequences of second-order cybernetics
Ingest: Evaluating Long-Context Reasoning in LLM-Based WebAgents (ingest-report) - NeurIPS workshop study injects irrelevant prior tasks into web-agent contexts and observes soft degradation, loops, and objective loss, extending controlled distractor findings to multi-session agents
Ingest: Every Good Regulator of a System Must Be a Model of That System (ingest-report) - Qualified good-regulator theorem vocabulary for models, goals, disturbances, and optimal regulation
Ingest: Everything you need to know about LLM memory (ingest-report) - Rosebud Journal memory essay reframing LLM memory as a policy stack over raw/derived artifacts, retrieval timing, curation, and forgetting propagation
Ingest: Evolution of the PDCA Cycle (ingest-report) - Historical evidence that PDCA/PDSA is the quality-engineering descendant of the conjecture-test learning loop
Ingest: Evolving Self-Reference: Matter, Symbols, and Semantic Closure (ingest-report) - Semantic-closure vocabulary and limits for using matter-symbol complementarity as a cross-representational analogy
Ingest: Externalization in LLM Agents (ingest-report) - Survey paper unifying LLM agent memory, skills, protocols, and harness engineering as externalized cognitive infrastructure rather than model-weight capability alone
Ingest: FALSIFYBENCH: Evaluating Inductive Reasoning in LLMs with Rule Discovery Games (ingest-report) - Wason 2-4-6 benchmark over 12 LLMs; lands as first behavioural evidence that falsification-seeking discriminates hypothesis-formers, and a process-scored counter-case to the known-target critique
Ingest: Foundation and History of the PDSA Cycle (ingest-report) - Moen's PDSA history as independent Deming-lineage corroboration that the KB's discovery-lifecycle learning core recurs outside philosophy of science
Ingest: From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence (ingest-report) - Epiplexity paper formalizing extractable structure for computationally bounded observers, useful for observer-relative information value and context-efficiency theory.
Ingest: From Human Memory to AI Memory: A Survey on Memory Mechanisms in the Era of LLMs (ingest-report) - Ingest of the 3D-8Q LLM-memory survey: how object/form/time taxonomy fits our memory-axes and cognitive-analogy-skepticism notes
Ingest: Frontis-MA1: Training an AI4AI Model toward Recursive Self-Improvement in MLE (ingest-report) - Frontis-MA1 couples execution-grounded operator training to fixed-controller evolutionary search, with large MLE gains but no revision of the improvement machinery
Ingest: Gentle-Coding (repo: README + Proof-of-Concept + RESEARCH) (ingest-report) - Gentle-Coding argues low-stakes prompt framing plus an explicit fallback Safety-Token cuts loops, freezing, and latency; three-snapshot repo capture (README, PoC, RESEARCH bibliography).
Ingest: Geometry of Knowledge Allows Extending Diversity Boundaries of Large Language Models (ingest-report) - Latent-conditioning framework raises LLM output diversity on NoveltyBench/AUT while authors admit a missing low-quality/OOD oracle — a clean positive instance of generate-cheap-verify-expensive at the embedding substrate.
Ingest: GIANTS: Generative Insight Anticipation from Scientific Literature (ingest-report) - GIANTS backcasts scientific discovery into a two-parent insight prediction benchmark, showing RL gains under a manufactured soft similarity oracle
Ingest: Graphiti: Temporal Knowledge Graph for AI Agents (ingest-report) - Graph-first agent memory with bi-temporal edge invalidation — the strongest counterexample to files-first architecture in the surveyed memory systems
Ingest: Gödel Machines — Provably Optimal Self-Improvements (ingest-report) - Schmidhuber's Gödel machine — rewrite of any part of one's own code, proof searcher included, gated on a proof of higher axiomatized utility: the proof-governed case of reflective self-modification
Ingest: Harness Engineering for Self-Improvement (ingest-report) - Harness self-improvement synthesis separates editable deployment machinery from the model, but its benchmark evidence remains bounded by fixed objectives, evaluators, and outer loops
Ingest: Harness Engineering Is Cybernetics (ingest-report) - Conceptual thread framing harness engineering as cybernetic feedback-loop design: sensors, actuators, constraints, and externalized judgment.
Ingest: Harness Engineering: Leveraging Codex in an Agent-First World (ingest-report) - Practitioner report on 1M LOC fully agent-generated codebase — harness engineering as constrain/inform/verify/correct, entropy management via background cleanup agents, error messages as dual-function constraining
Ingest: Harness Handbook (ingest-report) - Source-validated behavior maps improve harness edit-site localization, while executed modification and self-evolution remain unevaluated.
Ingest: Harness Updating Is Not Harness Benefit (ingest-report) - Primary cross-pairing evidence separates harness-edit production from artifact loading, faithful use, and downstream benefit, stopping short of compounding
Ingest: How I built a self-improving software factory (ingest-report) - A practitioner account of Fluent, a software factory that turns observations into tested, reviewed, merged changes and reusable project expertise.
Ingest: How Is LLM Reasoning Distracted by Irrelevant Context? (ingest-report) - GSM-DC isolates irrelevant-context effects and finds power-law error growth with distractor count, grounding soft context degradation while comparing training- and inference-time mitigations
Ingest: How to Build an AI Second Brain With Claude and Obsidian That Gets Smarter Every Day (ingest-report) - Convenient Claude+Obsidian second-brain setup guide -- useful adoption packaging, with live-account connectors as the risk boundary
Ingest: How to build your own agent harness??? (Mike Piccolo / iii) (ingest-report) - iii ships a production agent harness as independently-versioned workers on one bus behind a uniform trigger primitive, making harness layers hot-swappable
Ingest: How to Recursively Improve Your Agents (ingest-report) - Agno practitioner loop mines specifications and usage traces into probes, then iteratively edits a live agent within a fixed, weakly evaluated outer method
Ingest: How we built our knowledge base (ingest-report) - Cerebras practitioner report on an internal enterprise knowledge base: source-native ingestion, Slack/code retrieval, MCP tools, project scoping, and multi-consumer activation
Ingest: Human Bottlenecks (ingest-report) - Borretti argues AI value has a human-side competence floor — knowledge, executive function, and intelligence are bottlenecks software cannot lift.
Ingest: Human Routers of Machine Words (ingest-report) - Borretti polemic 'writing is thinking' — corroborating field evidence for reverse-compression and vibe-noting risks
Ingest: Huxley-Gödel Machine (ingest-report) - ICLR 2026 HGM paper arguing immediate benchmark score is a weak parent-selection signal for self-improving coding agents; clade-metaproductivity better predicts productive lineages
Ingest: HyperAgents (ingest-report) - HyperAgents paper provides cross-domain evidence for editable meta-agent transfer while leaving outer-loop machinery fixed and compounding unestablished
Ingest: Improving AI Skills with autoresearch & evals-skills (ingest-report) - Three-take Auto Research field report where optimization only worked after manual error analysis, failure taxonomy design, and judge calibration across the Three Gulfs.
Ingest: In Search of Lost Domain Generalization (ingest-report) - DomainBed shows tuned ERM matches specialized domain-generalization methods once model selection is declared, giving the reach-assessment cluster its first captured empirical counterweight
Ingest: in-toto — Providing farm-to-table guarantees for bits and bytes (ingest-report) - Cryptographic whole-chain supply-chain verification (in-toto) as a cross-domain exemplar for the KB's verification-cost, lineage, and staleness theory
Ingest: Infinite midwit (ingest-report) - Objective-vs-subjective intelligence essay arguing that AI's real bottleneck is taste and boringness judgment, not benchmarked competence
Ingest: Intelligent AI Delegation (ingest-report) - Google DeepMind delegation framework centers verifiability, liability, trust, and 11 task axes in agent delegation; notable for accountability vacuum and liability firebreaks in long chains
Ingest: Inter-language Reflection: A Conceptual Model and Its Implementation (ingest-report) - Mature model defining inter-language reflection as traditional reflection plus linguistic symbiosis through data and protocol mappings.
Ingest: Interpolation, Extrapolation, Hyperpolation (ingest-report) - Toby Ord's hyperpolation paper gives a geometric vocabulary for off-subspace creativity, sharpening the KB's synthesis-oracle and discovery/reach notes.
Ingest: Into the Unknown: Self-Learning Large Language Models (ingest-report) - Hallucination-driven self-learning LLM paper proposing Points in the Unknown, a self-question/search/train loop, and metrics for selecting models that can discover factual knowledge gaps
Ingest: Ironies of Automation (Bainbridge 1983) (ingest-report) - Bainbridge 1983 — foundational precursor whose monitoring/deskilling ironies supply period evidence for the KB's automation-boundary and effort-relocation notes
Ingest: Knowledge-Centric Self-Improvement (ingest-report) - Controlled knowledge-centric self-improvement protocol in which stateless agents curate a shared knowledge base through task forums, cross-task forums, and scoped distillation
Ingest: Language model harnesses are compositional generalizers (ingest-report) - Empirical RLM evidence that harness-level context offloading and programmatic sub-calls can turn length and domain shifts into locally familiar model calls.
Ingest: Language Models Don't Always Say What They Think (ingest-report) - Turpin et al. use controlled input biases to show chain-of-thought can rationalize changed answers while omitting the feature that caused the change
Ingest: Language Models Need Sleep (ingest-report) - Sleep operationalizes offline consolidation as upward self-distillation into slower model weights plus RL-guided dreaming, separating scheduling posture from representational form
Ingest: Language Models, Like Humans, Show Content Effects on Reasoning Tasks (ingest-report) - Empirical demonstration that LLMs mirror human content effects on reasoning (syllogisms, NLI, Wason) — content bias survives scaling and instruction tuning but chain-of-thought partially restores content-independent reasoning
Ingest: Language Symbiosis through Symbiotic Reflection (ingest-report) - Related 2002 version restating symbiotic reflection and up/down transfer; useful for lineage but not independent support.
Ingest: Large Language Model Agents Are Not Always Faithful Self-Evolvers (ingest-report) - Causal-intervention paper (v3, 13 backbones) showing self-evolving agents faithfully use raw trajectories but largely ignore condensed experience, making behavioral faithfulness the missing evaluation criterion for distilled memory
Ingest: Lessons from Building AI Agents for Financial Services (ingest-report) - Production financial-agent report supports S3-first files with derived PostgreSQL, skill shadowing for customization, and 'model eats scaffolding,' with fiscal-period normalization as a calculator counterexample
Ingest: Letta (MemGPT): Stateful Agents with Self-Managed Memory (ingest-report) - Agent memory platform where the LLM self-manages a three-tier memory hierarchy (core/recall/archival) using an OS analogy — the strongest existing exemplar of the agent-self-managed agency model, now evolving toward git-backed memory files
Ingest: LLM Knowledge Bases (ingest-report) - Karpathy on agent-maintained research wikis in Obsidian — index files and brief summaries replacing fancy RAG at roughly 100-article scale
Ingest: LLM Position Bias Benchmark (Swapped-Order Pairwise Judging) (ingest-report) - Swapped-order pairwise-judging benchmark showing that across 27 LLM judges the median model flips its underlying winner in 44.8% of decisive cases, with large model-dependent first-position lifts on sibling story-edit pairs
Ingest: LLM Wiki (ingest-report) - Karpathy's long-form agent-maintained wiki manifesto — explicit raw/wiki/schema architecture plus index/log separation beyond his earlier X-post workflow sketch
Ingest: Locating and Editing Factual Associations in GPT (ingest-report) - ROME localizes factual associations to mid-layer MLPs and edits them with a rank-one weight update — the nearest approach to our named explicit-retention falsifier, but it lacks scope-addressability
Ingest: Maximum Effective Context Window (ingest-report) - Study across 11 frontier LLMs finds maximum effective context windows up to 99% below advertised limits, task-dependent, with hallucinations approaching 100% beyond the effective window
Ingest: Mem0: Universal Memory Layer for AI Agents (ingest-report) - Mem0's two-phase add pipeline (extract facts + LLM-judged CRUD reconciliation) is the purest production example of automated accretion-without-synthesis — now contextualized by the comparative review
Ingest: Memento-Skills: Let Agents Design Agents (ingest-report) - Memento-Skills supplies empirical evidence for frozen-LLM deploy-time learning through routed, rewritten executable skill memory, with transfer bounded by domain alignment
Ingest: Memory Intelligence Agent (ingest-report) - MIA mixed-substrate deep-research agent memory paper — search trajectories become both workflow memory and Planner weight updates during test-time learning
Ingest: Memory Scaling for AI Agents (ingest-report) - Databricks memory-scaling experiments showing enterprise agent gains from external memory only when retrieval, distillation, and governance scale with the store
Ingest: Mesa Optimizers and Language Recursion (ingest-report) - Speculative essay arguing mesa optimizers may emerge suddenly because language recursion and learned search both compress many cases into reusable generative rules.
Ingest: Meta-Harness: End-to-End Optimization of Model Harnesses (ingest-report) - Controlled ablation showing raw execution traces (10 MTok/iter) outperform summaries by 10+ points in automated harness search — first empirical evidence for diagnostic richness as binding constraint
Ingest: Minimum Viable Ontology / Domain Maps (ingest-report) - Tweet thread proposing "minimum viable ontology" — the smallest term list to orient a newcomer in a domain — with a vibecoded prototype (domainmaps.co) and pedagogical framing via "conceptual thresholds"
Ingest: Multi-Agent Memory from a Computer Architecture Perspective (ingest-report) - Computer-architecture analogy for multi-agent memory — shared/distributed paradigms, three-layer hierarchy, consistency protocols as the critical unsolved problem
Ingest: Natural-Language Agent Harnesses (ingest-report) - NLAH paper externalizes agent control logic as portable natural-language artifacts — key empirical finding: explicit structure helps only when it tightens alignment with evaluator acceptance criteria, not by adding process layers
Ingest: Nested Learning (ingest-report) - Nested Learning recasts architectures and optimizers as nested associative memories, but Hope's broad gains vary local components within a fixed weight-only decomposition
Ingest: Novel Memory Forgetting Techniques for Autonomous AI Agents (ingest-report) - Formula-based adaptive forgetting with constrained optimization for agent memory — the inspectable alternative to RL-trained memory policy, with empirical evidence that uncontrolled accumulation causes false memory propagation
Ingest: On Doctors, Mechanics and Computer Specialists — Where are the Problems with Credence Goods? (ingest-report) - Credence-goods microeconomics paper — evidence and refinement for the verification-boundary claim, adding liability as an institutional substitute for verifiability
Ingest: On Learning How to Learn Learning Strategies (ingest-report) - Schmidhuber's reward-gated self-modification report as historical evidence for oracle-dependent behavior learning and reversible promotion
Ingest: On the "Induction Bias" in Sequence Models (ingest-report) - 190k-run empirical study showing transformers need orders-of-magnitude more data than RNNs for state tracking due to absence of step-by-step induction bias; introduces sharing factor kappa quantifying cross-length mechanism reuse
Ingest: Online Convex Programming and Generalized Infinitesimal Gradient Ascent (ingest-report) - Zinkevich's online convex-programming regret bounds provide a technical basis for direct, gateless evidence-responsive updates and their assumptions.
Ingest: Open continual learning that you fully control (ingest-report) - A short X design sketch that turns organizational expertise into evals, environments, agent behavior, and trace-driven continual improvement
Ingest: OpenClaw-RL: Train Any Agent Simply by Talking (ingest-report) - Deployment-time RL uses user replies, tool output, terminal feedback, and GUI state as next-state signals, collapsing training and deployment while challenging Commonplace's three-timescale model
Ingest: Orchestrate subagents at scale with dynamic workflows (ingest-report) - Claude Code dynamic-workflows docs — the saveable, script-authored counterpoint to RLM's ephemeral orchestrators; evidence for the orchestration/run-state persistence cluster
Ingest: Palantir Ontology vs Decision Traces (ingest-report) - Jaya Gupta frames Palantir-style top-down ontology and workflow-first decision traces as two ways to build LLM-facing world models
Ingest: Post by @deepfates — LLM "memory" as context stuffing (ingest-report) - Deepfates argues LLM "memory" is just context-stuffing that creates false salience (Chekov's gun), advocates agentic context-building, but concludes weight updates are necessary — directly contradicts this KB's durability-not-weights position
Ingest: Post by @koylanai (ingest-report) - Argues that pairwise judging plus round-robin win rates is a better evaluation primitive than absolute scoring for open-ended LLM tasks with no hard ground truth
Ingest: Professional Software Developers Don't Vibe, They Control (ingest-report) - Empirical study (N=112) finding experienced developers control AI agents through SE practices, not vibe coding -- grounds constraining, underspecification, and programming-practices-transfer arguments
Ingest: Prompt Stability in Code LLMs (ingest-report) - Code-LLM study finds performance and prompt stability are distinct, smaller models may be steadier, and emotional/personality variants expose confidence miscalibration missed by standard benchmarks
Ingest: PROV-Overview — An Overview of the PROV Family of Documents (ingest-report) - W3C PROV family roadmap — the canonical standard for 'full provenance' that the KB's lineage concept deliberately trims down from
Ingest: Psychology already solved AI memory — identity isn't stored, it's constructed (ingest-report) - Thread proposing five psychology principles (Conway, Damasio, Bruner, Klein & Nichols) for AI memory as identity construction — directly engages the KB's open question about whether cognitive science analogies are decorative or mechanistic
Ingest: Reading Between the Dots: Decoding Hidden Computation across Filler Tokens (ingest-report) - Filler-token study as evidence that monitorability depends on the observation surface and that hidden process traces can diagnose failures
Ingest: Recursive Language Models - what finally gave me the 'aha' moment (ingest-report) - Practitioner comparison of direct generation, RAG, ReAct, CodeAct, subagents, and RLM gives concrete evidence for REPL substrate, symbolic variable return, and scaffold-level truncation
Ingest: Reflection and Semantics in Lisp (ingest-report) - Foundational procedural-reflection vocabulary: embedded self-theory, two-way causal connection, reflective vantage point, and explicit-versus-absorbed processor state
Ingest: Researchers Asked LLMs for Strategic Advice. They Got "Trendslop" in Return. (ingest-report) - HBR trendslop article: LLM strategy advice follows fashionable management discourse despite prompt and context variation.
Ingest: Scaling Managed Agents: Decoupling the brain from the hands (ingest-report) - Anthropic Managed Agents report showing brain/hand/session interface decomposition, durable session logs, and stale harness assumptions as model capability changes
Ingest: Second-Order Cybernetics: An Historical Introduction (ingest-report) - Historical placement of observed versus observing systems and the observer's epistemological inclusion
Ingest: Self-Harness: Harnesses That Improve Themselves (ingest-report) - Same-model harness optimization improves three Terminal-Bench agents through structured failure mining and regression-gated local edits, within a tightly fixed outer method
Ingest: Self-Improving AI Coding Agents Through Accumulated Behavioral Rules (ingest-report) - Production evidence for review-derived coding-agent rules, bounded by an instance-to-rule oracle gap and an uncontrolled, fixed-architecture deployment.
Ingest: Self-Improving Algorithms (ingest-report) - Ailon et al. learn a product input distribution, then retain data structures that reach entropy-optimal limiting performance
Ingest: Self-Revising Discovery Systems for Science (ingest-report) - Category-theoretic discovery framework: typed-artifact copresheaves, retrieval/search/discovery typology, and MDL/AIC gates
Ingest: Self-training Large Language Models through Knowledge Detection (ingest-report) - EMNLP paper turning unknown-detection scores into filtered DPO preference data, with selective self-training reducing hallucination and limiting forgetting on Wikipedia QA
Ingest: Skill Synthesis — Materializing Knowledge as Skills (ingest-report) - Sentry co-founder's practitioner report on synthesizing Claude Code skills from domain-specific source material (commit history, security patches, OWASP docs) — found 8 real IDORs missed by professional pen testing
Ingest: SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning (ingest-report) - SkillRL paper showing trajectory-distilled skill banks co-evolving with GRPO-trained agent policy, bridging readable skill artifacts and weight-based learning
Ingest: Slate: Moving Beyond ReAct and RLM (ingest-report) - Practitioner architecture uses bounded worker threads that return compressed episodes to an orchestrator, converging on Commonplace's bounded-context model for memory, coherence, and decomposition
Ingest: Software Engineering of Self-Adaptive Systems: An Organised Tour and Future Challenges (ingest-report) - Weyns's six-wave engineering perspective supplies a bounded self-adaptation vocabulary and a warning that MAPE-K is an engineering reference model, not a membership definition.
Ingest: Solving a Million-Step LLM Task with Zero Errors (ingest-report) - MAKER achieves zero errors over one million LLM steps via maximal decomposition into single-step microagents with first-to-ahead-by-k voting and red-flagging — proves O(s ln s) cost scaling when hard per-step oracles exist
Ingest: Spacebot: AI Agent for Teams and Communities (ingest-report) - Spacebot README ingest covering process-typed concurrent agent runtime architecture, branch scoping, cortex supervision, and typed unified memory
Ingest: Structured Test-Time Scaling: From Multi-Agent Systems to General Inference Architectures (ingest-report) - Formal result connects topology compression, scope isolation, and verification as a causal chain enabling hierarchical multi-agent systems to avoid exponential error accumulation
Ingest: SuperARC — Can Complexity and Uncomputability Explain Intelligence? (ingest-report) - Ingest of SuperARC — AIT-grounded benchmark where frontier LLMs score phi ~0.03 while neuro-symbolic CTM/BDM achieves 1.000 on recursive compression; newer models regress; print-statement-only outputs demonstrate zero algorithmic abstraction
Ingest: Symbiotic Reflection between an Object-Oriented and a Logic Programming Language (ingest-report) - Primary precursor defining symbiotic reflection, mutual cross-language reasoning and action, and upping/downing entity transfer.
Ingest: Symbolic Learning Enables Self-Evolving Agents (ingest-report) - Early whole-harness optimizer treats prompts, tools, and pipeline topology as jointly learnable language-mediated artifacts, with cross-node credit assignment and same-oracle rollback
Ingest: The "Mismanaged Geniuses" Hypothesis (ingest-report) - Hypothesis that current frontier LMs are bottlenecked by learned decomposition/scaffold policy rather than base capability, using RLMs and orchestrator-subagent systems as evidence
Ingest: The Agent Loop Architecture (ingest-report) - Inngest practitioner framing of durable agent loops as loop + skill + orchestrator, useful for the run-state and skill-persistence boundary
Ingest: The Anatomy of a Design Theory (ingest-report) - Gregor and Jones's six-core/two-additional anatomy separates design-theory content from implementation agents and physical instantiations
Ingest: The Anatomy of an Agent Harness (ingest-report) - Practitioner taxonomy deriving harness components (filesystem, bash, sandboxes, memory, context management, long-horizon execution) from model limitations — provides the component anatomy that bridges Lopopolo's practice and the cybernetics framing
Ingest: The Best Model Routing is Task Specific (ingest-report) - Jerry Liu on task-specific model routing: generic gateways handle provider routing, but workflow-specific routers earn cost and accuracy gains from private evals and input taxonomies
Ingest: The Bitter Lesson (ingest-report) - Wikipedia-contextualized capture of Sutton's Bitter Lesson, useful for scaling arguments and caveats about general methods versus hand-coded knowledge.
Ingest: The Bug That Shipped (ingest-report) - 3,700-trial practitioner evidence that coding models can diagnose deployment failures when explicitly probed but rarely surface them in undirected self-review
Ingest: The File System Is the New Database: How I Built a Personal OS for AI Agents (ingest-report) - Practitioner report on a file-based personal OS for AI agents, useful as self-reported evidence for filesystem-first context engineering.
Ingest: The Flawed Ephemeral Software Hypothesis (ingest-report) - Essay distinguishing vibe coding from true software ephemerality, arguing that state, integration, interface stability, and auditability keep important systems anchored to durable artifact stacks.
Ingest: The Geometry of Forgetting (ingest-report) - Embedding-memory paper arguing that interference and low effective dimensionality, not time decay, drive forgetting and false recall in similarity retrieval.
Ingest: The GitHub for Context Doesn't Exist Yet (ingest-report) - Prukalpa on shared organizational context outliving agent-stack churn, with production failures motivating lineage, semantic impact review, controlled trace learning, security, and portability
Ingest: The Log Is the Agent (ingest-report) - Omnara essay arguing an agent IS its durable append-only event log; log ownership = agent ownership
Ingest: The Nature of Theory in Information Systems (ingest-report) - Gregor's mature taxonomy defines design-and-action theory by prescriptive purpose and exposes actionability as a separate theory-actor-context relation
Ingest: The No Free Lunch Theorem (ingest-report) - No Free Lunch explainer grounding codification as an unavoidable inductive-bias bet: strategies win only when their assumptions match the problem distribution.
Ingest: The Optimal Choice of Hypothesis Is the Weakest, Not the Shortest (ingest-report) - Bennett's extension-size alternative to minimum description length, bounded by its uniform task prior and fixed representational decomposition
Ingest: The Physics of Symbols: Bridging the Epistemic Cut (ingest-report) - Epistemic-cut vocabulary for symbolic constraints, material dynamics, and the limits of prose/code analogies
Ingest: The Price of Meaning: Why Every Semantic Memory System Forgets (ingest-report) - Formal no-escape theorem for semantic memory interference, with exact-record and symbolic-verifier escape clauses that sharpen retrieval-vs-verification tradeoffs.
Ingest: The Risks of Invariant Risk Minimization (ingest-report) - Formal bound on the causal route to reach-assessment: IRM recovers the invariant predictor only under narrow conditions, and admits solutions that look invariant on training and revert to ERM at test
Ingest: The Second Brain Trap (ingest-report) - PlugLab AI founder reframes "second brain" failure as stored knowledge that never activates in context, then proposes trigger-rich graph structure as the fix
Ingest: The Self-Healing Agent Harness (ingest-report) - Practitioner report where live agent-response graders feed tickets, auto-fixes, re-grading, and rollout gates instead of sitting as offline eval dashboards
Ingest: The Spec Is the New Code — A Guide to Spec Driven Development (ingest-report) - MercadoLibre engineering lead's practitioner guide to Spec Driven Development — the spec/plan/task/implement cascade as methodology for eliminating agent ambiguity, with ecosystem convergence evidence and maturity-level progression
Ingest: The Vision of Autonomic Computing (ingest-report) - Kephart & Chess 2003, origin of MAPE-K and the four self-* properties; the primary source behind the KB's reference-model-not-definition reading of self-adaptive loops.
Ingest: The What & When of Self-Evolving Agents (ingest-report) - 3×3 framework mapping self-evolving-agent updates across external files, harnesses, and model weights over task, session, and population horizons
Ingest: The Y-Combinator for LLMs (ingest-report) - λ-RLM preprint replacing open-ended RLM REPL code with typed combinators, formal bounds, and long-context benchmark evidence
Ingest: Thread by @frgx (ingest-report) - Figma's practitioner account of turning security precedents into a trusted, tested policy artifact that drives agent review, repo auditing, and secure code generation.
Ingest: Thread by @LechMazur (ingest-report) - Lech Mazur's public benchmark announcement compressing the headline position-bias result, the GPT-5.4 callout, and the operational motivation from everyday comparison prompts
Ingest: ToolGate: Contract-Grounded and Verified Tool Execution for LLMs (ingest-report) - ToolGate makes tool-state mutation transactional through typed symbolic state, pre-execution admissibility checks, and post-execution result verification
Ingest: Toulmin Argument (ingest-report) - Pedagogical treatment of Toulmin's six-part argument model — canonical source for the structured-claim type's Evidence/Reasoning/Caveats sections
Ingest: Towards a Science of AI Agent Reliability (ingest-report) - Reliability framework paper arguing mean task success is inadequate for agents, replacing it with consistency, robustness, predictability, and safety.
Ingest: Towards a Science of Scaling Agent Systems (ingest-report) - Controlled multi-agent scaling paper showing coordination gains depend on task decomposability, verification, and context overhead rather than agent count.
Ingest: Towards Automating Eval Engineering (ingest-report) - Eval Engineering Skill announcement describing trace mining, user-guided task design, Harbor environments, verifier inspection, and iterative agent improvement
Ingest: Towards Automating Scientific Review with Google's Paper Assistant Tool (ingest-report) - Google PAT paper as evidence for verifiable-subrole review automation: segmenting manuscripts, scaling inference, and keeping humans accountable for final review authority.
Ingest: Towards Causal Representation Learning (ingest-report) - Causal representation learning grounds the claim that causal models support intervention, counterfactual, and reusable-mechanism generalization
Ingest: Towards Faithfully Interpretable NLP Systems (ingest-report) - Jacovi and Goldberg separate faithfulness from plausibility, expose the assumptions behind common explanation tests, and argue for graded rather than binary faithfulness
Ingest: Towards Linguistic Symbiosis of an Object-Oriented and a Logic Programming Language (ingest-report) - Implementation evidence that linguistic symbiosis across logic and object-oriented paradigms requires explicit syntactic and semantic mappings.
Ingest: tracecraft (ingest-report) - S3-backed CLI coordination tool for multi-agent systems — exemplifies coordination-without-guarantees and the files-over-database bet applied to inter-agent state rather than knowledge storage
Ingest: Trajectory-Informed Memory Generation for Self-Improving Agent Systems (ingest-report) - IBM pipeline extracts strategy, recovery, and optimization tips from trajectories for runtime retrieval; subtask granularity yields +14.3-point AppWorld gains under a narrow task-completion oracle
Ingest: Transformers Learn In-Context by Gradient Descent (ingest-report) - Mechanistic ICML paper showing in-context regression can be implemented as gradient descent inside Transformer forward passes, sharpening the internal half of the KB's in-context-learning theory
Ingest: Verbalizable Representations Form a Global Workspace in Language Models (ingest-report) - Anthropic J-space paper as evidence for probeable parametric state, activation-vs-presence, and externalized reasoning as internal-workspace relief
Ingest: We Should Take Text Optimization More Seriously (ingest-report) - Manifesto arguing text optimization is a legitimate, sample-efficient update mechanism enabling update-time compute — upstream restatement of the KB continual-learning cluster
Ingest: What is an Agent Harness (ingest-report) - Arize taxonomy separates nine harness components and adds permission/safety plus lifecycle hooks as first-class primitives, a fifth independent convergence on the harness decomposition
Ingest: What spec-driven development gets wrong (ingest-report) - Augment's argument that spec-driven development fails unless agents co-maintain the spec — bidirectional spec as a mechanism for matching maintenance throughput to generation throughput
Ingest: What Survives in Multi-Agent Systems (ingest-report) - Applied bitter-lesson analysis predicting which multi-agent patterns survive stronger models — argues filesystem, forking, and spawning are structural while fixed orchestration is a vision feature
Ingest: When code is free, research is all that matters (ingest-report) - Investor/researcher argument that oracle availability (not capability) determines automation boundary for cognitive work — research taste is unautomatable because problem selection has no ground truth
Ingest: Where It Lives Is Not What It Is (architectural vocabulary for retained adaptation) (ingest-report) - External paper derived from Commonplace's four-field artifact analysis; the ingest tracks paper-from-notes lineage, adds sovereignty-risk refinements, and may become a citable authority after acceptance
Ingest: Where It Lives Is Not What It Is (June 2026 version) (ingest-report) - Updated self-authored ASISAS position paper adding 141-system corpus evidence to the four-field retained-artifact vocabulary.
Ingest: Why AI systems don't learn and what to do about it (ingest-report) - Position paper arguing current AI externalizes learning into human-run MLOps and proposing an A-B-M architecture where meta-control arbitrates observation and action learning for lifelong adaptation.
Ingest: Why Software Factories Fail (ingest-report) - Dex Horthy's failed lights-off software-factory case supports the verification boundary and challenges agent-only maintainability review
Ingest: Why Software Factories Fail: Benchmarking the new frontier (ingest-report) - Dex Horthy's SlopCodeBench run turns long-horizon code maintenance into a delayed benchmark signal while showing why deterministic proxies do not yet warrant lights-off autonomy
Ingest: Why Software Factories Fail: Turning the lights back on (ingest-report) - Dex Horthy's staged human-review workflow moves maintainability judgment upstream and limits unchecked agent work with vertical slices
Ingest: Your Old Agent Architecture Is Dead… Meet Its Replacement (ingest-report) - A strong architectural argument for recursive code–inference execution, plan-as-program workflows, and compiling natural-language guardrails into typed deterministic middleware.
Intelligent AI Delegation (snapshot) - Google DeepMind framework for intelligent AI delegation — proposes adaptive protocols covering task decomposition, multi-objective optimization, trust/reputation, verifiable completion, and security for human-AI and AI-AI delegation networks, with explicit analysis of how MCP, A2A, AP2, and UCP map onto these requirements.
Inter-language Reflection: A Conceptual Model and Its Implementation (snapshot) - Mature conceptual model of inter-language reflection through causal connection, data mappings, protocol mappings, and reflective access across languages.
Interpolation, Extrapolation, Hyperpolation: Generalising into new dimensions (snapshot) - Toby Ord's hyperpolation proposal - a third generalisation mode for reasoning outside the data subspace, linked to creativity and AI limits
Into the Unknown: Self-Learning Large Language Models (snapshot) - Self-learning LLM paper proposing Points in the Unknown, hallucination-based unknown detection, self-questioning/search/training loop, and self-learning capability metrics
Knowledge-Centric Self-Improvement (snapshot) - Controlled case studies of knowledge-centric self-improvement, where generic disposable agents curate a persistent shared knowledge base that transfers across tasks and LLM families.
Lab Leaks, Black Holes, and Eggs: Epistemic Case Study Competition (snapshot) - Future of Life Foundation competition brief for AI-assisted epistemic investigations and compounding knowledge bases, including its ingestion, structure, assessment, and personal-use implications.
Language model harnesses are compositional generalizers (snapshot) - Zhang and Khattab's empirical argument that context offloading and programmatic sub-calls let RLM harnesses generalize across task length and domain by keeping root-model observations locally in-distribution.
Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought (snapshot) - Controlled biasing experiments show chain-of-thought can rationalize answers while omitting the input feature that causally shifted them
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories (snapshot) - Sleep-paradigm paper on upward self-distillation into growing low-frequency memory blocks and RL-guided dreaming for continual language-model learning
Language Models, Like Humans, Show Content Effects on Reasoning Tasks (snapshot)
Language Symbiosis through Symbiotic Reflection (snapshot) - Related 2002 version of symbiotic reflection across logic and object-oriented languages, with mutual reasoning, modification, and automatic up/down entity transfer.
Large Language Model Agents Are Not Always Faithful Self-Evolvers (snapshot) - Causal-intervention paper showing self-evolving agents rely on raw trajectories more faithfully than condensed experience, exposing a compression-faithfulness gap across frameworks, models, and environments
Lessons from Building AI Agents for Financial Services (snapshot)
Letta (MemGPT): Stateful Agents with Self-Managed Memory
LLM Wiki (snapshot) - Karpathy's idea file for agent-maintained personal wikis, centered on a persistent markdown layer between raw sources and query-time chat
Locating and Editing Factual Associations in GPT (snapshot) - ROME paper — factual associations in GPT localize to mid-layer feed-forward modules and can be directly edited via rank-one weight updates, evidence that factual knowledge is stored as identifiable, editable computation rather than diffuse distributed representation.
Mem0: Universal Memory Layer for AI Agents (snapshot)
Memento-Skills: Let Agents Design Agents (snapshot) - Memento-Skills continually learns without weight updates by routing, rewriting, and expanding persistent structured-Markdown skills.
Memory Intelligence Agent (snapshot) - MIA paper on converting deep-research search trajectories into workflow memory and Planner test-time training
Memory Scaling for AI Agents (snapshot) - Databricks AI Research argument and experiments for external-memory scaling as a third agent improvement axis alongside model and inference scaling
Mesa Optimizers and Language Recursion (snapshot) - Speculative blog post connecting mesa optimizers to language recursion by treating both as compressed generative rules that can appear as sudden capability jumps.
Meta V-JEPA 2 action-conditioned world model (source-review) - Meta's V-JEPA 2 article and paper page ground action-conditioned world models as predictors for planning in new physical environments
Meta V-JEPA world-model framing (source-review) - Meta's V-JEPA article frames LeCun's JEPA line as latent predictive world modeling for generalized reasoning and planning
Meta-Harness: End-to-End Optimization of Model Harnesses (snapshot) - Stanford/MIT paper proposing Meta-Harness, an outer-loop system that uses a coding agent with full filesystem access to prior code and execution traces to automatically search over and optimize LLM harnesses — outperforming hand-engineered baselines on text classification and TerminalBench-2.
Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead (snapshot) - Position paper reframing multi-agent memory management through a computer architecture lens — proposes shared vs. distributed memory paradigms, a three-layer hierarchy (I/O, cache, memory), and identifies memory consistency as the most urgent unresolved challenge for scalable multi-agent systems.
Natural-Language Agent Harnesses (snapshot) - Proposes externalizing agent control logic (contracts, roles, stages, failure taxonomy) as portable natural-language artifacts (NLAHs) with an Intelligent Harness Runtime, evaluated on SWE-bench and OSWorld — key finding: explicit structure helps only when it tightens alignment with evaluator acceptance criteria.
Nested Learning: The Illusion of Deep Learning Architecture (snapshot) - Nested Learning paper recasting architectures and optimizers as multi-level associative memories, with Hope experiments spanning continual learning and long-context reasoning
Novel Memory Forgetting Techniques for Autonomous AI Agents: Balancing Relevance and Efficiency (snapshot) - Adaptive budgeted forgetting framework for long-horizon conversational agents — relevance scoring (recency, frequency, semantic alignment) plus constrained optimization to prune memory while reducing false memory propagation.
On Doctors, Mechanics and Computer Specialists Or Where are the Problems with Credence Goods? (snapshot) - Unifying game-theoretic model of credence goods showing that most conflicting results in the literature (Darby-Karni, Pitchik-Schotter, Wolinsky, Emons, Taylor) reduce to relaxing one of three conditions (homogeneous consumers, verifiability-or-liability, commitment) that otherwise yield a costless efficient market.
On Learning How to Learn Learning Strategies (snapshot) - Schmidhuber's 1995 technical report introducing incremental self-improvement through reward-gated self-modification of learning strategies
On the "Induction Bias" in Sequence Models (snapshot)
On the Criteria To Be Used in Decomposing Systems into Modules (snapshot) - Two working modularizations of the same KWIC index program, indistinguishable by running them, distinguished only by whether module boundaries hide the design decisions likely to change — the information-hiding criterion.
Online Convex Programming and Generalized Infinitesimal Gradient Ascent (snapshot) - Carnegie Mellon technical report introducing online convex programming, regret guarantees, and generalized infinitesimal gradient ascent (GIGA).
OpenClaw-RL: Train Any Agent Simply by Talking (snapshot) - Framework that converts live next-state signals (user replies, tool outputs, terminal feedback, GUI state) into RL rewards and token-level supervision, enabling a single policy to personalize and improve on agentic tasks simultaneously.
Orchestrate subagents at scale with dynamic workflows (snapshot) - Claude Code's dynamic workflows feature docs — JavaScript scripts that orchestrate up to 1,000 subagents per run, covering when to use them, how to write/save/manage them, and how they compare to subagents and agent teams.
Post by @deepfates (snapshot)
Post by @JayaGup10 (snapshot)
Post by @karpathy (snapshot)
Post by @koylanai (snapshot)
Post by @Vtrivedy10 (snapshot)
Professional Software Developers Don't Vibe, They Control: AI Agent Use for Coding in 2025 (snapshot)
Prompt Stability in Code LLMs (snapshot)
PROV-Overview: An Overview of the PROV Family of Documents (snapshot) - W3C roadmap document to the PROV family of specifications for representing and interchanging provenance information on the web.
Psychology already solved AI memory — identity isn't stored, it's constructed (snapshot) - Thread arguing AI memory should adopt psychology's model of identity construction through autobiographical memory, citing Conway, Damasio, Bruner, and Klein & Nichols
Reading Between the Dots: Decoding Hidden Computation across Filler Tokens (snapshot) - Mechanistic study showing frontier open-weight models perform causally decodable computation across content-free filler tokens despite having no readable chain of thought
Recursive Language Models - what finally gave me the 'aha' moment (snapshot)
Reflection and Semantics in Lisp (snapshot) - Brian Cantwell Smith's foundational procedural-reflection account: embedded self-theory, bidirectional causal connection, vantage point, and 3-Lisp's reflective tower
Researchers Asked LLMs for Strategic Advice. They Got “Trendslop” in Return. (snapshot) - HBR article naming strategy trendslop: LLMs favor fashionable management options even when prompts and business context change.
Scaling Managed Agents: Decoupling the brain from the hands (snapshot) - Anthropic's Managed Agents architecture argues for stable brain, hand, and session interfaces that outlast changing agent harness implementations.
Second-order cybernetics: an historical introduction (snapshot) - Scott's historical reconstruction of second-order cybernetics as the study of observing systems
Self-Harness: Harnesses That Improve Themselves (snapshot) - Self-improvement loop that mines model-specific trace failures, proposes bounded harness edits, and regression-gates them on Terminal-Bench
Self-Improving AI Coding Agents Through Accumulated Behavioral Rules: A Closed-Loop Framework (snapshot) - Production case study of converting accepted code-review feedback into persistent, version-controlled behavioral rules shared across coding-agent interfaces.
Self-Improving Algorithms (snapshot) - Ailon et al.'s self-improving algorithms that learn an unknown product input distribution and converge to entropy-optimal sorting and Delaunay-triangulation performance
Self-Revising Discovery Systems for Science (snapshot) - Category-theoretic framework for agentic scientific discovery systems that models artifact states as copresheaves and distinguishes retrieval, search, and discovery as structurally different operations — grounded in a materials-science case study.
Self-training Large Language Models through Knowledge Detection (snapshot) - EMNLP 2024 paper on self-training LLMs by filtering DPO preference data to unknown samples using reference-free contradiction scores
Skill Synthesis: Materializing Knowledge as Skills (snapshot)
SkillOpt: Executive Strategy for Self-Evolving Agent Skills (snapshot) - SkillOpt paper treating agent skill documents as trainable external state with bounded text edits, validation gates, rejected-edit buffers, and slow/meta updates
SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning (snapshot) - SkillRL paper on distilling successful and failed agent trajectories into a hierarchical skill bank co-evolved with GRPO policy training
Software Engineering of Self-Adaptive Systems: An Organised Tour and Future Challenges (snapshot) - Full-text capture of Weyns's six-wave engineering perspective on self-adaptive systems, including its conceptual model and future challenges.
Solving a Million-Step LLM Task with Zero Errors (snapshot)
Spacebot: AI Agent for Teams and Communities (snapshot)
Structured Test-Time Scaling: From Multi-Agent Systems to General Inference Architectures (snapshot) - Unified theoretical framework explaining how three structural mechanisms (topology compression, scope isolation, verification) enable hierarchical multi-agent systems to bypass exponential error accumulation in test-time scaling.
SuperARC: Recursive Compression Benchmark (snapshot) - Introduces SuperARC, an AIT-grounded benchmark showing frontier LLMs score near zero on recursive compression tasks and newer versions often regress, while neuro-symbolic CTM/BDM methods achieve perfect scores — evidence that statistical pattern matching differs fundamentally from algorithmic abstraction.
Symbiotic Reflection between an Object-Oriented and a Logic Programming Language (snapshot) - Early formulation of symbiotic reflection across logic and object-oriented languages, including the upping/downing schema for entity transfer.
Symbolic Learning Enables Self-Evolving Agents (snapshot) - Agent symbolic learning treats prompts, tools, and their composition as jointly optimizable language-based weights in a self-evolving agent network.
The "Mismanaged Geniuses" Hypothesis (snapshot)
The Agent Loop Architecture (snapshot)
The Anatomy of an Agent Harness (snapshot)
The Best Model Routing is Task Specific (snapshot)
The Bitter Lesson (snapshot)
The Bitter Lesson (Sutton, 2019) — original essay (snapshot)
The Bug That Shipped (snapshot)
The File System Is the New Database: How I Built a Personal OS for AI Agents (snapshot)
The Flawed Ephemeral Software Hypothesis (snapshot) - Essay arguing AI makes software more malleable, not ephemeral, because validation, state, interface stability, and auditability remain the load-bearing bottlenecks.
The Geometry of Forgetting (snapshot) - Embedding-space account of human-like forgetting and false memory — interference, low effective dimensionality, and semantic clustering reproduce classic memory effects.
The GitHub for Context Doesn’t Exist Yet (snapshot)
The Log Is the Agent (snapshot)
The No Free Lunch Theorem: Why No Learning Algorithm Is Universally Best (snapshot) - Bruce Nielson's No Free Lunch explainer connecting inductive bias, optimization, neural-network smoothness assumptions, and Popperian fallibility
The Optimal Choice of Hypothesis Is the Weakest, Not the Shortest (snapshot) - Formal and experimental argument that extension size, rather than description length, predicts generalisation under a uniform task prior
The Physics of Symbols: Bridging the Epistemic Cut (snapshot) - Pattee's rate-independent symbol and rate-dependent dynamics account of the epistemic cut
The Price of Meaning: Why Every Semantic Memory System Forgets (snapshot) - Formal no-escape theorem paper arguing semantic memory systems face interference-driven forgetting and false recall under finite effective dimensionality.
The Risks of Invariant Risk Minimization (snapshot) - First formal analysis of the IRM objective — shows it can fail catastrophically unless test data are sufficiently similar to training data, and offers no gain over ERM in non-linear regimes
The Second Brain Trap (snapshot) - PlugLab AI article arguing that second-brain systems fail when stored notes never activate in working context
The Self-Healing Agent Harness (snapshot)
The Spec Is the New Code. A Guide to Spec Driven Development (snapshot)
The What & When of Self-Evolving Agents (snapshot) - A 3×3 framework classifying self-evolving agents by update substrate—external files, agent harness, or model weights—and persistence horizon.
The Y-Combinator for LLMs: Solving Long-Context Rot with λ-Calculus (snapshot) - λ-RLM paper replacing open-ended recursive LLM control code with typed λ-calculus combinators for bounded long-context reasoning.
Thread by @frgx (snapshot)
Thread by @LechMazur (snapshot)
Thread by @melodyskim (snapshot)
ToolGate: Contract-Grounded and Verified Tool Execution for LLMs (snapshot) - Contract-grounded tool execution gates invocation and state updates with typed symbolic state plus Hoare-style preconditions and postconditions.
Toulmin Argument (snapshot)
Towards a Science of AI Agent Reliability (snapshot)
Towards a Science of Scaling Agent Systems (snapshot)
Towards Automating Eval Engineering (snapshot)
Towards Automating Scientific Review with Google's Paper Assistant Tool (snapshot) - Google Paper Assistant Tool paper on inference-scaled AI review, STOC/ICML author pilots, and roles for AI automation in peer review.
Towards Causal Representation Learning (snapshot) - Schoelkopf et al. review how causal models, interventions, and reusable mechanisms bear on transfer, robustness, and out-of-distribution generalization
Towards Faithfully Interpretable NLP Systems: How Should We Define and Evaluate Faithfulness? (snapshot) - Jacovi and Goldberg separate faithfulness from human-plausible explanation and argue for graded, assumption-explicit evaluation
Towards Linguistic Symbiosis of an Object-Oriented and a Logic Programming Language (snapshot) - Implementation study of linguistic symbiosis between object-oriented Smalltalk and logic-programming SOUL across differing paradigms.
tracecraft (snapshot) - S3-backed CLI coordination layer for multi-agent AI systems — shared memory, messaging, task claiming, and barriers stored as JSON files in any S3-compatible bucket, with no servers or databases required.
Trajectory-Informed Memory Generation for Self-Improving Agent Systems (snapshot) - IBM Research framework that extracts three categories of actionable tips (strategy, recovery, optimization) from agent execution trajectories and injects them at runtime — evaluated on AppWorld showing up to 14.3 pp gains in scenario goal completion.
Transformers Learn In-Context by Gradient Descent (snapshot) - Mechanistic ICML paper showing linear self-attention can implement gradient descent for in-context regression and trained Transformers can recover that construction
Untitled (snapshot) - Moen and Norman's concise correction of common myths about the Deming cycle and the historical distinction between PDCA and PDSA.
Untitled (snapshot) - Gregor distinguishes five theory types and elaborates theory for design and action as normative guidance whose articulated principles concern artifacts and their construction.
Untitled (snapshot) - Moen and Norman's historical account of the PDCA-to-PDSA lineage and its development into the Model for Improvement.
Untitled (snapshot) - Moen's source history of Deming's PDSA cycle from early scientific method through Shewhart, PDCA, and Deming's 1993 formulation.
Untitled (snapshot) - Bainbridge's foundational account of how automation can intensify operator problems and make retained human skill harder to sustain.
Untitled (snapshot) - Gregor and Jones specify six core and two additional components for information-systems design theory, including scope, mutability, justification, implementation, and instantiation.
Untitled (snapshot) - Gregor classifies information-systems theories by primary goal and defines design-and-action theory through explicit prescriptions for constructing artifacts.
Untitled (snapshot) - Kephart and Chess's original autonomic-computing vision, including the managed-element and autonomic-manager architecture behind MAPE-K.
Verbalizable Representations Form a Global Workspace in Language Models (snapshot) - Anthropic interpretability paper arguing that verbalizable J-space representations act as a limited global workspace for LLM reasoning, auditing, and training interventions
We Should Take Text Optimization More Seriously (snapshot)
What is an Agent Harness (snapshot)
What spec-driven development gets wrong (snapshot)
What Survives in Multi-Agent Systems
When code is free, research is all that matters (snapshot)
Where It Lives Is Not What It Is: Vocabulary for Retained Adaptation in Agentic Systems (snapshot) - Position paper proposing a four-field architectural vocabulary (storage substrate, representational form, lineage, behavioral authority) for classifying retained behavior-shaping artifacts in agentic systems, beyond storage-first labels.
Where It Lives Is Not What It Is: Vocabulary for Retained Adaptation in Agentic Systems (snapshot) - Updated position paper proposing a four-field architectural vocabulary for retained behavior-shaping artifacts in agentic systems and applying it to a 141-system agent-memory corpus.
Why AI systems don't learn and what to do about it (snapshot) - Dupoux, LeCun, and Malik argue current AI externalizes learning into human-run MLOps, then propose an A-B-M architecture where observation learning, action learning, and a meta-control plane are integrated for lifelong adaptation.
Why Software Factories Fail (snapshot)
Why Software Factories Fail: Benchmarking the new frontier (snapshot)
Why Software Factories Fail: Turning the lights back on (snapshot)
Writing conventions for kb/sources/ (descriptive profile)
Your Old Agent Architecture Is Dead… Meet Its Replacement (snapshot)
ШІ-агенти: чотири навички, які важливіші за добрий промпт (snapshot) - Mezha article applying Anthropic's AI Fluency framework to agent use: delegation, description, discernment, and diligence as operator skills beyond prompting.

Keys	Action
`?`	Open this help
`n`	Next page
`p`	Previous page
`s`	Search