Skip to main content
Back to timeline
arXivSource publication:

An agent that evolves without touching model weights: Harness-RSI keeps accepted prompt, memory and code revisions for the next run

Synopsis

Harness-RSI puts an agent's self-evolution outside the model weights: prompts, memory, workflows and executable components can be revised, accepted revisions persist and load in later runs, and the task model stays fixed. It separates update targets from proposal and admission mechanisms, characterizes search as evolutionary optimization and reliability as continual PAC learning, and under explicit assumptions yields an expected generalization bound, a retention bound conditional on historical regression budgets, and a high-probability admission rule using fresh independent validation data. On SpreadsheetBench, GPT-5.4-nano with HarnessOpt and skill reaches 77.58%, above GPT-5.5 with SkillOpt at 76.20%.

Source-provided article image: Self-Evolving Agent Harness
Figure 1 · arXiv · Page 1

Interpretation

An agent can improve outside its model: the harness of prompts, memory, workflows and executable components is revised while model weights stay fixed, and accepted revisions persist across runs. Earlier self-improvement routes either changed model weights or used scaffolds with no update mechanism; Harness-RSI treats the harness itself as the editable object and requires feedback-driven updates, update continuity and persistence across runs to hold together. On SpreadsheetBench, GPT-5.4-nano with HarnessOpt and skill reaches 77.58%, exceeding GPT-5.5 with SkillOpt at 76.20%; across all four models, HarnessOpt with skill improves on SkillOpt-Lite by 5.69 to 11.39 percentage points.

The framework splits what gets edited from how candidates are proposed and admitted, so revisions can be classified, inspected and rolled back. Existing methods differ in what they modify, how candidates are generated and how revisions are admitted, and empirical selection alone cannot establish generalization or retention. Update targets are divided into interpreted state and executable revisions, the latter into component implementations, repository-level control logic and plugin composition; the analysis derives an expected generalization bound, a retention bound conditional on historical regression budgets, and a high-probability admission rule using fresh independent validation data under explicit assumptions.

Plugin organization lets an agent generate, validate and mount new tools during execution and keep them for later tasks. A fixed tool set can only use capabilities registered in advance, whereas plugin architectures extend the editable object to the bindings that decide which implementation supplies a runtime service. In a simulated office workload, after Codex creates batch-lookup and summarization tools, first-task tool calls fall from 13 to 3 while wall time rises 54.9% and tokens rise 15.7%; a separately reported cumulative series shows savings from the second task, reaching 110,351 tokens and 22.97 s after five tasks.

Perspective

The scope is harness self-evolution with model weights held fixed, covering prompts, memory, workflows, component implementations, repository-level control logic and plugin composition; weight-only adaptation and scaffolds without an update mechanism fall outside it. The SpreadsheetBench comparison and the simulated office tool-generation case are the concrete settings reported, and reliable deployment requires continual validation as task distributions and evaluators change, with independent safeguards for learning, runtime privileges and consequential actions.

The theoretical results rest on explicit assumptions, so the generalization bound, retention bound and admission rule are not unconditional; in the plugin case, first-task cost increases coexist with cumulative savings, and the two records differ unresolvedly in first-task tokens and time, so they cannot be treated as one cost trajectory.