Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents
The paper argues that agent control logic can be learned into reusable code instead of rebuilt inside every prompt.
The authors introduce “Growing Harness,” which starts from a strategy-free scaffold and uses task failures to edit the harness itself. Its repair loop localizes failures with execution traces, batches fixes across a window of failures, and rolls back edits that hurt held-out successes. In tests on BrowseComp-Plus and WebArena-Verified, it led five of six benchmark-model settings and sharply reduced LLM calls and inference cost versus a Tool-Calling agent. ArXiv · AI/CL/LG's note
The authors introduce “Growing Harness,” which starts from a strategy-free scaffold and uses task failures to edit the harness itself. Its repair loop localizes failures with execution traces, batches fixes across a window of failures, and rolls back edits that hurt held-out successes. In tests on BrowseComp-Plus and WebArena-Verified, it led five of six benchmark-model settings and sharply reduced LLM calls and inference cost versus a Tool-Calling agent. ArXiv · AI/CL/LG's note
score 6