AI agents improve when their surrounding software is automatically redesigned

With models held fixed, the authors’ automated search raised average scores across six benchmarks by 10.89 percentage points over initial agent software.

Big Tech

Zhi Hong · Qianlong Wang · Hongchen Wei · Aidan Wu · Daohong Ying · Mingze Kong · +2 more

South China University of Technology · Fudan University · Wuhan University · Cornell University · University of California, San Diego

Research Digest·5 November 2026·2 min read

Illustration: AI agents improve when their surrounding software is automatically redesigned

In one line

HarnessFlow achieves the highest mean test scores on six benchmarks by using MCTS-guided search to optimize LLM agent harnesses.

Before you read this

Three ideas this paper builds on

agent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.165 papers →monte carlo tree searchA search algorithm that uses random sampling to evaluate actions and builds a tree for planning.5 papers →coding agentAn AI-powered software system that autonomously writes, modifies, or extends source code based on high-level instructions.144 papers →

Read it as

Large language models (LLMs) are often turned into agents that perform tasks by interacting with tools. The code that manages these interactions, the harness, is usually designed by hand, which is slow and costly. Testing a single harness requires running the full agent, making optimization expensive.

HarnessFlow treats the harness as code that an AI can edit. It uses a search algorithm called that balances exploration and improvement. A coding AI then refines the harness using past results and reusable components.

On six standard tests, HarnessFlow achieved the highest average scores. It improved performance by 10.89 percentage points over the baseline and by 3.95 percentage points over a prior automated method, while also reducing compute costs. This shows that guided search can automate harness design, reducing the need for manual engineering.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

agent harness monte carlo tree search coding agent harness evolution cost-quality trade-off self-evolving agent