Daily AI CatchupMiles v0.1: Open System for Post-Training AI Agents at Scale
LMSYS released Miles v0.1, an open system for improving AI agents through reinforcement learning after initial training. Miles handles rollout, sandboxing, asynchronous training, replay, model updates, and multi-hardware orchestration—enabling teams to run continuous improvement loops at scale. For example, coding agents can attempt tasks in isolated environments, collect successful results, and push updated models to workers without pipeline interruption.
Why it matters
💻 Developer · Miles abstracts away the infrastructure for distributed post-training. Use it to build self-improving agents without managing rollout, sandboxing, and training orchestration yourself.
📦 Product · Post-training on real user tasks means your agent gets incrementally better as it runs. Miles makes this feasible at scale without dedicated infrastructure teams.
🎨 Design · Agents trained on real failures improve in ways base models don't. Design for feedback loops: let users flag agent errors so they feed back into continuous improvement.
📈 Business · Continuous post-training compresses time-to-product. Instead of waiting for the next model release, you improve your deployed agent by learning from live tasks.
🤔 Just Curious · This democratizes post-training, which was previously locked behind org-scale infrastructure. Now smaller teams can run the same improvement loops that major labs use.