AAI News Hub
ResearchMon, August 17, 2026·2d ago2 sources corroborating

Researchers target RL training for agent harnesses

LEGO-RL and ClawGym II propose sandboxed, proxy-based methods for training long-horizon agents.

Why it matters

The work addresses a practical gap between agent harnesses used at inference time and policy-gradient training systems. Better harness-native RL methods could make long-horizon agent training more stable, observable and closer to real deployment conditions.

The key points

  • 1.LEGO-RL targets coding-agent harnesses with proxying and sandbox orchestration.
  • 2.ClawGym II frames harness training as black-box RL over captured calls.
  • 3.Both papers emphasize training-inference consistency for long-horizon agents.

Two new arXiv papers describe reinforcement learning frameworks for agents that operate through complex harnesses used in long-horizon tasks such as coding. LEGO-RL focuses on coding-agent harnesses, using in-process LLM proxying, sandbox orchestration, validation and monitoring tools to align training with native harness execution. ClawGym II presents a black-box RL framework that isolates rollouts in temporary sandboxes, captures model calls at the serving boundary, reconstructs multi-turn trajectories with prefix trees, and adapts PPO and GRPO for the recovered structure.

Try this today

Read the papers before designing RL pipelines for coding or long-horizon agents that depend on harness execution.

Sources & original reporting

This brief summarizes and links to reporting from the publishers below.

Enjoyed this brief? Get the next one in your inbox.

More in Research