Research Editor selected
Meta$^2$: Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents
Meta$^2$ is a new method that precomputes reusable knowledge for evolving agent harnesses (prompts, tools, memory, hooks) online during long-horizon tasks. It learns a textual Meta-Skill offline by contrasting training rollouts with and without privileged verifier information, then applies this knowledge online to decide whether to keep or evolve the harness per task. On Terminal-Bench 2.1 and DeepSWE 1.1, it achieves 81.00% and 44.44% with GPT-5.6-Terra, outperforming baselines.
Why it matters: This paper presents a novel approach to harness optimization that combines offline learning with online adaptation, showing strong benchmark results and addressing a key limitation of existing methods.
agent harness optimizationmeta-skill learningonline evolution