Meta's EvoHarness-RL got Qwen3-8B to 96.9% on ALFWorld.

The method turns tool use from a hardcoded prompt into a learned runtime behavior, then applies cost-aware RL teaching the agent when reading external state is worth the token budget. Qwen3-8B reaches a 96.9% average success rate against SkillOS at 80.2% and SkillRL at 89.9%, within two points of Claude Opus 4.5, and the same harness pushes Opus 4.5 to 98.5%. The claim builders should test rather than accept: harness design, not model size, is where the remaining points live. (VentureBeat)