AIDE2 boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This analysis characterizes the agent-harness improvement mechanism presented in Recursive self-improvement of AI research agents. Its intended use is to explain the documented design, context-management mechanisms and evaluation limits, not to certify an executable implementation. The target is in scope because model-driven research agents propose harness rewrites and evaluate the resulting agents. It is a builder or improvement plane, not merely a memory library: context management is one editable component of the research agent (SRC-1, captured paper §§1–2 and §3.5).

In the inner loop, a research agent optimizes code against a measurable objective on problems drawn from a diverse set of AI R&D tasks. The outer loop performs a meta-level optimization over this inner-loop research process, rewriting the research agent with the objective of improving the research efficiency of the inner-loop agent. --- /home/zby/llm/commonplace/.commonplace/worktrees/aide2-6b3bcd4791cf/kb/agentic-system-analyses/state/AAS-2026-10-14-2609-26457-6b3bcd4791cf-01/sources/arxiv-2609.26457v1-2026-10-14.txt

The functional inclusions are the documented two-loop search; generation and review of task solutions; aggregation of private evaluation scores; incumbent selection and reuse of accepted agent code; and the discovered agents' search, prompt compaction, failure-memory and robustness mechanisms. The original and human-engineered agents belong to the account in their documented roles as initial agent, outer-loop driver and comparison baseline. The reported selection trajectory, held-out evaluations, ignition test and appendices are also included, but only as author-reported design and operation (SRC-1, §§2–3, §5, appendices A–E).

The agent then returns one solution of its own choosing, which we denote x^t\hat{x}_{t}. Which candidate to edit next, which to return, and how to structure the agent’s memory are all part of the agent’s code and are editable by the outer loop. --- /home/zby/llm/commonplace/.commonplace/worktrees/aide2-6b3bcd4791cf/kb/agentic-system-analyses/state/AAS-2026-10-14-2609-26457-6b3bcd4791cf-01/sources/arxiv-2609.26457v1-2026-10-14.txt

External dependencies include model APIs, task codebases and public/private scoring procedures, and execution infrastructure for benchmark solutions. The paper names different models for the two loops and CPU/GPU infrastructure for held-out evaluations; those descriptions establish dependencies, not independently inspected provider or infrastructure behavior (SRC-1, §2.2 and appendix A). Model training, provider internals, benchmark implementations and the wider Weco production product are excluded. Their exclusion prevents conclusions about model-weight self-improvement, evaluator integrity beyond the paper's account, deployment guarantees or the full production agent.

The boundary kind is complete artifact, partial loop. The complete paper text is frozen, but no repository or executable checkout was supplied or acquired. Repository-tree coverage and shipped entry-point/consumer tracing therefore cannot be performed within this source. This is not a whole-system implementation boundary. The paper describes state production, acceptance and subsequent consumption, but does not provide inspectable code paths, full prompts, persistence/reload contracts or raw execution traces in this capture. Those gaps prevent verification of actual state lifecycles, runtime enforcement and independent reproduction. They do not prevent a bounded account of the documented mechanism.

The immutable source is arxiv-2609.26457v1-2026-10-14, a UTF-8 text capture derived from https://arxiv.org/html/2609.26457v1. HTML entities were decoded, whitespace normalized, and scripts/styles omitted. The capture header records the original HTML digest and extraction limits. The frontmatter SHA-256 pins the exact resulting capture bytes. The applicability cutoff is 2026-10-14, the supplied run date; the captured paper identifies itself as v1 dated 22 September 2026. No later revision, live page, linked repository or cited paper is part of the evidence boundary. The overall evidence tier is doc-grounded.

Source register