AI automatic optimization execution framework makes Haiku 4.5 terminal performance number one: Lin Junyang says this is exactly the "environment design" turning point he predicted

According to 1M AI News monitoring, researchers from Stanford, MIT, and the Korean game company KRAFTON released Meta-Harness, a methodology for an AI-driven execution framework that automatically optimizes performance (a harness, i.e., a wrapper model and an execution scaffold that drives agent actions, covering prompt design, tool calls, and context management). Unlike manually written execution frameworks, Meta-Harness has a coding agent read the code, execution logs, and scores of prior candidate frameworks to automatically iterate and improve them.

On the TerminalBench-2 terminal operation benchmark, Meta-Harness brought Claude Haiku 4.5’s success rate to 37.6%, surpassing Goose (35.5%) and Claude Code (27.5%), ranking first among all reported Haiku 4.5 execution frameworks. On Claude Opus 4.6, the success rate was 76.4%, ranking second.

Lin Junyang, the former technical lead of Qianwen (Tongyi), forwarded the post from the paper’s authors and commented: “ ‘Model + execution framework’ has already surpassed ‘just the model.’ An agent’s performance will be significantly influenced by the design and quality of the framework, and I truly believe this is the right direction.” In a long post he published on March 27 (now deleted), Lin Junyang had already predicted that environment design would evolve from a side project into a true startup category. Meta-Harness backed up this judgment with experimental data: using the same model, switching to an AI-optimized execution framework could produce performance differences of up to 10 percentage points.