Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents

Source facts

Growing Harness learns the agent harness itself from a strategy-free scaffold. Function-level traces localize failures, an optimizer repairs bounded code surfaces, and a held-out success-first gate rolls back repairs that harm prior capability. The authors report fewer LLM calls and lower deployed-agent inference cost across BrowseComp-Plus and WebArena-Verified.

Analysis

This work is a direct example of moving recurring control decisions out of model context and into persistent executable code. The resulting specialist agents may reduce inference cost and allow smaller models to perform stable workflows, provided rollback and evaluation gates are trustworthy.

Uncertainty/Doubt

• This is an arXiv preprint; benchmark gains, cost reductions, and transfer across model scales were not independently reproduced here.