AI Agents Gain Autonomy with EvoHarness-RL

A new framework called EvoHarness-RL is enhancing AI agent autonomy by teaching them to dynamically manage their external state. Developed by Meta AI and the University of Illinois Urbana-Champaign, this approach moves beyond predefined rules, allowing agents to learn when and how to use tools and information. This is crucial for complex, long-horizon tasks where an agent must track progress, recover from errors, and adapt to changing environments.

EvoHarness-RL consolidates an agent's support systems into a unified Belief, Progress, and Experience (BPE) workspace. This interface enables agents to track the live environment (Belief), manage task completion (Progress), and reuse past knowledge (Experience) through simple meta-actions. A two-stage training process, including cost-aware reinforcement learning, teaches agents efficiency in accessing this external state.

Evaluations on the ALFWorld benchmark demonstrated significant performance gains. A Qwen3-8B model trained with EvoHarness-RL achieved a 96.9% success rate, notably outperforming its baseline and even matching the performance of larger, more expensive models like Claude Opus 4.5. The framework also showed benefits for existing frontier models, improving their execution rates substantially.

The system exhibits dynamic adaptation, with agents learning to reduce reliance on external tools for routine tasks ('harness annealing') while scaling up their use for novel situations ('harness evolution'). This intelligent tool usage translates to reduced latency and compute costs for enterprises. EvoHarness-RL can integrate with existing systems, offering a flexible state-management layer without requiring complete overhauls.