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

Summary

The paper introduces Growing Harness, a training paradigm in which an LLM agent's orchestration code — the harness — is itself learned from task feedback. Training starts from a 'strategy-free' scaffold that exposes fixed model and tool interfaces but contains no controller. Failed executions are traced at function level; an offline optimizer model (GPT-5.6-terra or GPT-5.4) synthesizes repair candidates constrained to traced functions and an edit budget; each candidate must fix enough failures and pass a success-first gate on 50 held-out tasks, otherwise the repair sequence is transactionally rolled back. Across BrowseComp-Plus and WebArena-Verified (200 train / 50 gate / 50 final tasks) with gpt-oss-120b, gpt-oss-20b and Qwen3.5-4B, the method attains the highest mean success in five of six benchmark-model settings and trails by 0.7 pp in the sixth (39.3 vs 40.0 on BrowseComp-Plus with gpt-oss-20b), while cutting LLM calls 76.0-91.8% and deployed-agent cost 74.4-98.6% versus a Tool-Calling agent. On WebArena-Verified, success stays 44.7-45.3% across model scales while Tool-Calling drops to 6.7% with the 4B model. Main limitations: 50-task evaluations with confidence half-widths of ±11-13 pp and no significance tests, single-run ablations, no comparison against cited learning-based optimizers such as AutoHarness or Meta-Harness, unreported offline optimization cost, and no code release.