'Harness Handbook' (154 Upvotes) Targets the Exact Bottleneck in Self-Modifying Agent Harnesses: Behavior Localization

arXiv / HuggingFace Daily Papers·high signal

The paper argues that as an agent harness evolves — handling prompts, state, tool invocation, and execution coordination — 'behavior localization is therefore a central bottleneck in harness evolution': you can't safely edit what you can't find. It proposes the Harness Handbook, an automated behavior-centric map built from static analysis plus LLM assistance that links each behavior to its source, and Behavior-Guided Progressive Disclosure (BGPD), which walks agents down a behavioral hierarchy while verifying candidate code locations. Reported gains in planning efficiency for scattered code sites and cross-module interactions, at fewer tokens. Directly applicable to anyone running a self-improving harness.