TRACE: Train Agents on Their Own Failures With GRPO

Your RL-trained agent plateaus. Here's the fix Stanford just shipped.

On July 13, 2026 the Scaling Intelligence Lab at Stanford released TRACE — Turning Recurrent Agent failures into Capability-targeted training Environments. It went straight to the top of the ML feeds, and for a good reason: it fixes the most frustrating thing about training LLM agents with reinforcement learning.