Meta-Harness AI Auto-Optimization Framework Helps Haiku4.5 Win Terminal Test; Lin Junyang Optimistic About Environment Design Startup Direction

Researchers from Stanford, MIT, and South Korean game company KRAFTON have launched Meta-Harness, a method for AI auto-optimized execution frameworks. It uses a coding agent to read code, logs, and scores of candidate frameworks for automatic iterative optimization, replacing manually written frameworks (including prompt design, tool calls, and context management).

In the TerminalBench-2 benchmark test, Meta-Harness increased Claude Haiku4.5's pass rate to 37.6%, ranking first among all Haiku4.5 frameworks; Claude Opus4.6's pass rate reached 76.4%, ranking second.

Lin Junyang, former technical lead of Tongyi Qianwen, stated that "model + framework" has surpassed pure models, and agent performance is significantly affected by framework design quality. He predicted environment design would become a startup category, and Meta-Harness's data confirms this: the performance gap of the same model with an optimized framework can reach 10 percentage points.