Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents
Growing Harness converts recurring agent control into reusable code, cutting LLM calls by 76 to 92 percent.
Growing Harness is a failure-guided method that starts from a strategy-free scaffold and accumulates accepted code edits into one shared agent harness, moving recurring control out of the model context. Across BrowseComp-Plus and WebArena-Verified with deployment models from 4B to 120B parameters, it posted the highest mean success in five of six settings and trailed the best mean by 0.7 percentage points in the sixth. Relative to a tool-calling agent, it cut LLM calls by 76.0-91.8% and inference cost by 74.4-98.6%. On WebArena-Verified, success held at 44.7-45.3% across scales, while tool-calling fell to 6.7% with the 4B model.
- Failure-guided edits grow one shared harness from a strategy-free scaffold.
- Highest mean success in five of six BrowseComp-Plus and WebArena settings.
- LLM calls drop 76.0-91.8% and inference cost 74.4-98.6% versus tool-calling.
- WebArena success stays 44.7-45.3% even with a 4B deployment model.
From the original article
1 min read at arxiv.orgLarge language model (LLM) agents often handle streams of related tasks, yet standard harnesses repeatedly ask the model to reconstruct the same control decisions inside each task's context. We study whether task feedback can instead turn recurring control into reusable executable code, while reserving LLM calls for task-specific semantic reasoning.
By Laizhen Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.