Meta-Harness: Automatic Agent Harness Optimization

Source: https://yoonholee.com/meta-harness/ Author: Yoon Ho Lee et al. Date: 2025

Summary

System for automatically optimizing agent harnesses (prompts, tool definitions, control logic) by giving the optimizer access to full source code, scores, and execution traces of every prior candidate — 10M tokens of diagnostic context per step. Ranks #2 among Claude Opus agents on TerminalBench-2 (76.4% pass rate).

Key Claims

  • Novel approach: optimizer gets 10M tokens of diagnostic context (vs 26K for competing methods)
  • Full access to source code, scores, execution traces of all prior candidates — not just summaries
  • Text classification: 48.6% accuracy vs 40.9% for prior work, using 4x fewer context tokens
  • Math reasoning: +4.7 percentage points across 5 held-out models from a single retrieved strategy
  • TerminalBench-2: #2 Claude Opus (76.4%), #1 Haiku (37.6%)
  • Key insight: diagnostic richness enables more targeted optimization than aggregate metrics

Connection to Other Sources

Directly relevant to Anthropic’s agent eval framework and effective harnesses. Meta-Harness automates what those articles describe as manual harness design.

Concepts