h t t p s : / / a r x i v . o r g / a b s / 2 6 0 8 . 2 3 2 8 3 v 1

Ingest: Apodex 1.1: Scaling Agentic Intelligence for Complex Work

Type: kb/sources/types/ingest-report.md

Classification

This is a scientific preprint and system technical report: it defines a task formalism, presents model and harness designs, and reports benchmark and process-evaluation results. Author: the Apodex Team, which built the model and released the matching FrontierAgent repository and model weights; this gives the report direct design knowledge but also a strong interest in favorable framing.

Summary

The report defines working capability as sustained progress through executable state toward a verifiable delivery, then attributes Apodex 1.1 to two coupled scaling programs: broader file, search, and code environments, and learned coordination over decomposition, delegation, asynchronous returns, verification, and replanning. A shared harness and Apodex AgentOS hold task, workspace, provenance, budget, and publication state; SFT and agentic RL allegedly promote environment and coordination traces into model policy; ReAct, Agent Team, benchmark suites, HDS6 process judging, and three cases supply the evaluation. For Commonplace, the report is most useful as a detailed design case for active-work state, runtime-layer separation, coordination guarantees, fixed-decomposition analysis, and process-aware evaluation. The released code confirms several runtime mechanisms, but it does not reproduce the training or scores and materially weakens the paper's claims about Task Board semantics, default planning and verification, evidence-graph synthesis, and stale-output reconciliation, so the source should guide design questions rather than establish comparative performance or component causality.

Code Grounding

The reviewed implementation is the full-SHA FrontierAgent revision; its README links the matching report and describes the runtime, terminal product, and evaluation suite. The generic runtime implements Stateful ReAct and Agent Team workflows. AgentBus provides asynchronous submission, collection, and cancellation through asyncio-backed jobs and structured results, while SpawnGuard constrains depth, parallelism, token use, time, and resource release (AgentBus; SpawnGuard). Live intervention is also concrete: a queued inbox wakes aggregation, messages enter at safe loop boundaries, and admission renews a wall-time lease (steering queue; tool-loop injection; lease renewal).

The runtime has a run-scoped in-process Task Board, optional planning restrictions, and a finalization check for unresolved work (board state; planning transitions; finalization gate). The paper, however, says each board row carries dependencies and returned evidence or artifact references; the actual row schema contains description, resolution, owners, and group, with neither field family (schema declaration). The independent-verifier condition is a planning-mode gate keyed by verifier session name, and all three shipped Agent Team profiles disable planning mode, so planning before delegation and mandatory independent verification are configurable behavior, not default invariants (verifier gate; profile defaults).

Context and execution controls are substantially implemented. Tiered compaction uses provider-reported prompt or input-token usage, first removes older tool-result bodies while protecting fan-in results, and calls an LLM summarizer only when cheaper relief is insufficient (trigger and first tier; escalation and recovery). Soft wall-time budgeting and best-effort final recovery are present in the budget implementation, and file and shell tools enforce the documented /inputs, /workspace, and /outputs namespace policy (framework documentation). Output publication has a single authorized owner, an exact manifest, fail-closed non-publisher writes, and quota-bounded shared scratch space (publisher assignment; manifest enforcement; deliverable policy). Static inspection did not locate the paper's additional baseline-at-lease-grant and terminal-reconciliation mechanism, so the stronger claim that stale or empty files cannot count as delivery remains paper-only.

The synthesis and evaluation evidence is mixed. The released reporter is an optional, fail-open two-call path that reviews captured URL and snippet candidates and projects canonical citation inputs before writing (fast reporter). It records an empty per-subagent DAG and says the heavier ReAct reporter is not included (reporter boundary; empty DAG), which supports only a weaker form of evidence-aware synthesis than the paper's dedicated claim-evidence graph and writer stage. A public evaluation harness registers benchmark adapters and judges and runs isolated, timeout-bounded, resumable per-question subprocesses with configurable concurrency (harness overview; runner), but no retained result traces or complete artifacts tied to the paper's tables were located. No implementation was found for the reported environment registry and construction pipeline, the 33-domain/318-occupation/1,208-cluster inventory, replay generation, model-soup SFT, agentic RL, or PIVOT-RL.

This grounding used static source inspection only. No project dependencies, model weights, datasets, training jobs, tests, benchmark runs, or end-to-end harness executions were run. Inspection can establish mechanisms and defaults at the pinned revision; it cannot reproduce performance, throughput, quality, training effects, or causal attribution.

Quotes

  • Source extract (verbatim): Delegation, staged return, synthesis, branch revision, and recovery are represented in the training trajectories rather than treated only as properties of an inference-time wrapper.
  • Source location: Section 3.4 ("Training"), opening paragraph
  • Source extract (verbatim): To balance specialization with general capability, we train SFT variants over major capability domains, including general, agentic, and coding data, and combine them through model-soup merging. The resulting checkpoint serves as the unified behavioral initialization for subsequent optimization stages.
  • Source location: Section 3.4.1 ("Supervised Fine-Tuning"), closing paragraph
  • Source extract (verbatim): PIVOT-RL uses hindsight-guided trajectory localization over a large policy-training corpus.
  • Source location: Section 3.4.2 ("Reinforcement Learning"), paragraph "PIVOT-RL: Localized Optimization at Consequential Decisions"
  • Source extract (verbatim): At each pivot, we preserve the useful prefix and construct a localized continuation task with a short corrective hint. The hint provides directional guidance for the local correction, is never a prediction target, and is absent at inference time; for stateful tasks, we also restore the corresponding executable environment state.
  • Source location: Section 3.4.2 ("Reinforcement Learning"), paragraph following "PIVOT-RL: Localized Optimization at Consequential Decisions"
  • Source extract (verbatim): Apodex 1.1 externalizes this architecture onto a persistent task board and adds four capabilities: asynchronous human intervention, asymmetric verification, adaptive Max Team Effort, and evidence-grounded synthesis.
  • Source location: Section 3.2 ("Apodex Agent Team 1.1: Interactive Self-Organizing Teams"), opening paragraph
  • Source extract (verbatim): It is the shared coordination record between the model, runtime, and user: subagents receive their scope from it, completed work is attached back to it, and plan revisions are expressed as tool-mediated edits to it.
  • Source location: Section 3.2.1 ("From Latent Decomposition to an Explicit Task Board")
  • Source extract (verbatim): The coordination plane is also run-scoped rather than a durable distributed database. The current task board and Agent Bus sessions live in the worker process, and the runtime does not atomically checkpoint them together with the workspace filesystem.
  • Source location: Section 3.3, "Operational Boundaries" discussion closing the AgentOS runtime section (p. 17)

Connections Found

This source is a code-qualified technical basis, and in places a limitation, for active work state as distinct from chat history and for separating scheduling, context, and external state. Its external Task Board, asynchronous execution state, compacted model history, and filesystem namespaces instantiate those layers, while the paper's own non-atomic recovery boundary and the code's narrower board schema show where their coordination remains incomplete.

Apodex is also a useful compound design point for the multi-dimensional orchestration design space and coordination guarantees beyond channels: it combines learned delegation, an external board, asynchronous fan-in, resource guards, targeted verification roles, and single-owner publication. The implementation qualification matters because some advertised guarantees are optional or absent rather than default-enforced.

HDS6 is a concrete precedent for not inferring path health from final task success and a useful comparison for the trajectory-aware evaluation proposal. Its outcome-blind, evidence-citing roles and integrity gate inspect the execution route, but the report supplies neither a blinded output-only arm nor repeat-run calibration. Finally, the training and benchmark interpretation rests on learning inside a fixed decomposition inheriting its mistakes: trajectories, board and workspace state, verifier feedback, and localized decision windows can condition a model policy that composes tool use, delegation, collection, verification, recovery, and publication, but task contracts, environment and action families, board protocol, runtime, verifier and reporter designs, publication rules, budgets, datasets, and evaluation remain outside the effective update space. The reported improvement therefore supports the compound model-plus-harness configuration, not the necessity or superiority of those fixed choices.

Extractable Value

  1. A recovery boundary must co-version control state and mutable executable state. Apodex explicitly leaves its run-scoped Task Board and Agent Bus outside an atomic filesystem checkpoint; compared with ScienceFlow's workspace restoration, this supports a new design claim that scheduler ownership and dependencies must resume at the same logical point as the workspace while evidence and spent budget remain monotonic. [deep-dive]
  2. Active-work state can be made operational without treating the transcript as authoritative. The implemented board, execution registry, safe-boundary interventions, and finalization checks provide a concrete runtime case for the active-work-state note, while the schema mismatch prevents importing the paper's richer dependency and evidence fields as implemented facts. [quick-win]
  3. A configurable guard is not a deployed invariant. Planning-before-delegation and independent verification exist behind planning mode, yet shipped team profiles disable that mode; this is a reusable code-grounding test for claims about orchestration guarantees and defaults. [quick-win]
  4. Process evaluation needs a discriminating control, not only a richer trace judge. HDS6 contributes specialized evidence-citing roles and an integrity gate, but an output-only blinded arm, repeated-run calibration, and controlled trace mutations are still needed to learn whether trajectory access improves discrimination. [experiment]
  5. Agent-learning results need an explicit fixed-decomposition ledger. Here the learner can update a model policy over rich trajectories and composable runtime operations, while the task contract, environment families, action surface, coordination protocol, verifiers, publication policy, budgets, and evaluation stay fixed; this makes the model-plus-harness system the supported unit and blocks component-level causal claims. [deep-dive]
  6. Exact-manifest publication and stale-output reconciliation are separate guarantees. The repository confirms the former but not the paper's claimed baseline comparison at termination, providing a compact counterexample to treating access control as proof of fresh delivery. [quick-win]

Limitations (our opinion)

The report is an author-produced arXiv preprint about the authors' own model and product. Its large benchmark tables mix reported external values, internal reproductions, different harnesses, and proprietary systems with incomplete parameter information. The released repository contains benchmark machinery but not the result traces or complete artifacts needed to reproduce the tables, and this inspection did not run any benchmark. The performance rankings, throughput, model-scale efficiency, and causal attribution of gains therefore remain author-reported.

The central intervention is compound. ReAct versus Agent Team changes scaffolding and organized compute, while the release and report bundle model training, environment data, runtime mechanisms, task protocols, and verification choices. No component-isolating ablations establish that the Task Board, asynchronous fan-in, asymmetric verification, evidence synthesis, publication controls, environment construction, or PIVOT-RL caused the reported gains. The learner-visible signals include executable-environment trajectories, coordination traces, board and workspace state, verifier feedback, and localized decision windows; its operations include fixed tools, delegation, assignment, collection, verification, and publication; and its hypothesis class is the model policy over those choices. Everything else in the harness and evaluation is fixed. Improvement within that policy space cannot validate the fixed decomposition or show that omitted representations, actions, or protocols would be worse.

Several paper claims are stronger than the public implementation. Board rows lack the claimed dependencies and returned evidence or artifact references. Planning and independent-verifier gates are disabled in shipped profiles. The released reporter is optional and fail-open, carries no per-subagent evidence DAG, and omits the heavier backend. Exact-manifest output control is present, but stale-output reconciliation was not located. Static source inspection cannot decide whether private production code supplies these mechanisms, but the public revision cannot be cited as implementing them.

HDS6 is a promising process-assessment design, not evidence that access to a trajectory improves judging. It lacks a paired output-only control and reported independent calibration; one comparison also contrasts eight aggregated Apodex 1.0 runs with one Apodex 1.1 run. The curated cases show detailed workflows but cannot substitute for retained benchmark traces, and one case explicitly lacks its agent-level trace. These boundaries limit both evaluation conclusions and transfer to Commonplace review design.

Draft a note titled Reliable resume requires a checkpoint boundary across control and executable state that synthesizes Apodex's non-atomic Task Board/Agent Bus/filesystem boundary with ScienceFlow's restoration design, while keeping evidence and spent resource accounting outside rollback.