What's this about?
Agent Lightning v1.0 (16 upvotes) starts from the observation that modern agents operate inside harnesses managing tools, context and control flow — making the harness a critical part of the system. It uses a disaggregated architecture that connects reinforcement learning training to agents running in their real harness rather than in a simplified training environment.
What exactly is the problem with training outside the harness?
Reinforcement learning needs the agent to act and be scored, and the simplest setup strips everything down: model, task, reward. But the deployed agent doesn't work that way — it has retries, a summarizer, a control loop, tool timeouts. Training in the stripped environment optimizes for a system that doesn't exist, and the resulting behaviours may be wrong or irrelevant once the harness is back. Training in the real harness is more faithful and much harder to scale, because the harness is slow, stateful and not differentiable. "Disaggregated" means separating the training machinery from the agent execution so each can run where it fits.
Why is this worth noticing?
- It's the fourth distinct harness-focused approach in a fortnight — evolve it (DarwinX), engineer it (StateM), train through it black-box (ClawGym II), and now train inside it with proper infrastructure.
- The train/deploy mismatch it targets is an old lesson from robotics, where policies learned in simulation famously fail on real hardware. The harness is the agent equivalent of the sim-to-real gap.