Meta-Harness beats Claude Code on TerminalBench-2

This is a strong reminder that for agentic systems, the harness can matter almost as much as the base model, and raw traces beat summary compression when failures are subtle.

• –The paper reports filesystem access to prior code, scores, and logs, with as much as 10M tokens of diagnostic context per step.
• –On TerminalBench-2, Meta-Harness reports 76.4% on Claude Opus 4.6 versus 58.0 for Claude Code, and 37.6% on Claude Haiku 4.5 versus 27.5 for Claude Code.
• –The same framework also improves text classification by 7.7 points over ACE while using 4x fewer context tokens, and lifts math reasoning by 4.7 points on average across five held-out models.
• –The interesting twist is that Claude Code is both part of the optimization loop and a benchmark entry, so the result is more about search and instrumentation than a simple model-vs-model bake-off.