NTH

HarnessFlow: MCTS-Guided Harness Optimization for LLM Agents

AuthorsZhi Hong, Qianlong Wang, Hongchen Wei, Aidan Wu, Daohong Ying, Mingze Kong, Bei Liu, Chong Luo

AffiliationsSouth China University of Technology Fudan University · Wuhan University Cornell University University of California, San Diego · The Chinese University of Hong Kong, Shenzhen Microsoft Research Asia∗ Corresponding author. ‡ Work done during an internship at MSRA

November 17, 2026 2 min read
Watch on YouTube
The one-line take

HarnessFlow uses tree search and a coding agent to automatically refine the software harnesses that guide LLM agents.

Key results

69.48%
Average test score

Mean across six benchmarks after search.

10.89
Gain over initial harness

Percentage-point improvement in average test score.

3.95
Gain over Meta-Harness

Percentage-point improvement in average test score.

15
Refinement attempts

Candidate attempts per benchmark.

74.2%
Best validation score by round 15

Mean best-so-far validation score.

0.232
Average execution cost

USD per task, averaged across benchmarks.

What the paper found

HarnessFlow treats an LLM agent’s complete executable harness—not just its prompt or workflow—as the object to optimize. With the model and task environment held fixed, it stores candidate harnesses, scores, edits, and execution traces in a refinement tree; Monte Carlo Tree Search balances refining high-scoring designs against exploring underused branches, while a coding agent draws on that history and a reusable component library to modify context construction, state, control flow, and tool use. Tested with GPT-5.4-mini on TAU2-bench, AppWorld, WebShop, BrowseComp-Plus, SWE-bench Verified, and BigCodeBench-Hard, HarnessFlow achieved a 69.48% average test score after 15 refinement attempts per benchmark—10.89 percentage points above the initial harness and 3.95 points above Meta-Harness. Its best-so-far validation score reached 71.5% after three candidate evaluations and 74.2% by round 15. The system also compared favorably with hand-engineered agents including Anthropic’s Claude Code and OpenAI’s Codex; its average execution cost was $0.232 per task, lower than those baselines. The central contribution is a budget-aware search process that preserves and revisits complete harness designs instead of relying on a single sequence of edits.

Original abstract

The performance of large language model (LLM) agents depends on both the underlying model and the harness that governs interaction with tools and the environment. Yet harness design remains largely manual, and automating it requires navigating an open-ended search space through costly end-to-end evaluations. We introduce HarnessFlow, a framework for optimizing complete executable harnesses under a limited candidate-evaluation budget, with the model and task environment held fixed. HarnessFlow organizes candidates and their evaluation histories in a refinement tree, using a Monte Carlo Tree Search (MCTS)-based policy to balance exploitation and exploration. A coding agent refines selected harnesses using archived code, execution feedback, and a library of reusable harness mechanisms. With 15 refinement attempts per benchmark, HarnessFlow achieves the highest mean test scores among the compared methods on six benchmarks spanning agentic tasks, coding, and web interaction. The optimized harnesses achieve an average score of 69.48%, outperforming the initial harnesses and Meta-Harness by 10.89 and 3.95 percentage points, respectively, while incurring lower average per-task execution costs than hand-engineered baselines. These results suggest that structured harness search offers a promising route toward automating the design of capable and efficient LLM agents.

Read the original paper

More in AI Agents

Browse all 77 papers →