Researchers propose unified framework for agent harnesses that rewrite their own code
An arXiv paper describes a design framework in which an agent's runtime software diagnoses its own failures, patches its code, and validates updates before deployment.
In plain words
An agent harness is the software layer around an AI model that manages memory, tools, and checking of results. The paper proposes a framework where this layer improves itself: it studies its own execution logs, rewrites parts of its code, and tests the changes before use. The authors report this worked better than prompt tweaks or full rewrites on a coding benchmark.
What was announced
The paper argues that whether a foundation model succeeds as an autonomous agent depends as much on its runtime software, the agent harness, as on the model's weights. A harness tuned for one model or task often stalls, wastes tokens, or regresses on another, which the authors say has pushed development from prompt tuning toward harnesses that inspect their own execution traces, separate runtime bugs from model limits, and rewrite their own code.
The framework pairs four editable runtime components — context and state, agent control, action interfaces, and verification and recovery — with a four-stage loop that diagnoses failures, proposes code patches, validates candidates, and deploys updates. The authors report that modular updates with validation gates generalize across models while bounding token cost, whereas full-harness rewrites tend to overfit or break across modules. In experiments on SWE-bench Lite, they say gated modular evolution outperformed prompt reflection and full-harness rewrites across models and repositories at a fraction of the token cost. They also describe defenses against failure modes including prompt bloat, benchmark gaming, and evaluator tampering.
What it does not show
- The source is only the paper's abstract; the full paper's methods, experimental setup, and results have not been reviewed, and no independent confirmation of the reported findings is available.
For security and IT teams
- Scope
- The framework and its SWE-bench Lite results are claims from the paper's authors, reported in the abstract; the excerpt does not cover implementation details or availability.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.