Meta's EvoHarness-RL: Agents Learn to Autonomously Manage External Frameworks
omarsar0 · x · 2026-08-10
Meta has published new research on AI agents titled EvoHarness-RL. Current agent harnesses are mostly hand-authored, making it difficult to tune robust behaviors for long-horizon tasks.
The proposed method allows agents to learn harness policies offline and dynamically construct or update external states online during runtime. Specifically, the model learns the action space via supervised harness fine-tuning, followed by cost-aware GRPO to explore when to read, update, and consolidate during long-running tasks. Experiments show that Qwen3-8B achieves a 96.9% success rate on the ALFWorld benchmark.
Two key dynamics emerge from the training process:
- Harness annealing: Recurring harness-use patterns get absorbed into the model's own policy, shifting the agent from frequent calls to selective access.
- Harness evolution: Progress updates and experience consolidation are effectively compressed.
More from coding & agent
- DIRECT: retrieval-based next-question suggestions for conversational agents — burkov · 2026-10-16
- Claude Code and Codex now autocomplete the user's next action — tokenbender · 2026-10-16
- Hermes Agent ships official Reverse Engineer Anything plugin with one-command setup — Teknium · 2026-10-16
- MLExpert: 35B MoE Agent Trained to Autonomously Handle Machine Learning Engineering — Shangheng Du · 2026-10-16
- Cowork-Gym: Alibaba Synthesizes 64.9K Agent Work Environments, Boosting Qwen by 8.7 Points — Anni Zou · 2026-10-16
- MMSearch-OPD: Joint RL and On-Policy Distillation for Long-Horizon Multimodal Search Agents — PekingUniversity · 2026-10-16