Meta researchers taught an 8B AI model to match Claude Opus 4.5 — without the frontier price tag

Meta AI and University of Illinois Urbana–Champaign researchers introduced EvoHarness-RL, a framework that trains AI agents to decide when to read, update, and consolidate information from their runtime environment. The approach aims to replace rigid developer-written harness rules with learned behavior, helping smaller models handle long-horizon enterprise workflows more reliably.

The work suggests that better agent infrastructure and training can narrow the gap between compact open models and expensive frontier systems. If validated broadly, agent harnesses could become a key optimization layer, letting enterprises improve reliability and cost efficiency without always upgrading to larger models.