Self-Harness Lets Agents Modify Their Own Wrappers
量子位 · wechat · 2026-07-19
The Shanghai Artificial Intelligence Laboratory team has proposed Self-Harness. Instead of altering the underlying model, the model analyzes its own execution trajectories to identify failure modes and then proposes constrained modifications to the outer Harness. Regression tests determine whether these changes are adopted.
Tested on Terminal-Bench-2.0 across multiple backend models, this approach—keeping the base model, tool environment, and evaluation protocol identical while only modifying the Harness—yielded significant improvements: a 104% total gain for Qwen3.5-35B-A3B, 28% for MiniMaxM2.5, and 24% for GLM-5.
The paper breaks down the three-step pipeline: weakness mining, Harness proposal, and regression verification. It provides examples showing that different models exhibit distinct typical issues, such as delaying file output, looping endlessly after tool failures, or struggling to manage shell states and the transition from exploration to implementation.
Related event: Agent Harness Self-Improvement and Domain-Specific Design(6 posts)→
More from coding & agent
- Kolab Benchmark Uses Pro Screen Recordings to Show AI Agents Overestimate Collaboration Quality — minnesotanlp · 2026-10-27
- Dev argues Claude burns far more tokens than GPT models for the same coding work — haider1 · 2026-10-27
- Stop building agents that can't change models: stable contracts, external state, real evals — rseroter · 2026-10-27
- Cheap models can cost more in agent harnesses than frontier ones, researcher warns — vasuman · 2026-10-27
- Kanzo: A 5k-Line Rust Kanban Binary With MCP for Claude Code and Codex — allmudi · 2026-10-27
- 'A human approves everything' is a trap: approval fatigue and how to actually design safe agents — Efistoffeles · 2026-10-27