What's this about?
TRACE is a method to trace an agent's final outcome back to the specific steps in its "trajectory" (its sequence of actions and the context it saw), so the system can automatically figure out what context helped or hurt — and improve what it feeds the agent next time.
What exactly is "context engineering"?
Context engineering is the craft of deciding exactly what information to put into a model's limited context window — which files, memories, tool outputs, and instructions — to get good results. Done by hand it's fiddly; TRACE tries to automate it by attributing success/failure to particular pieces of context.
Why is this notable?
As agents run longer, multi-step tasks, managing their context becomes the bottleneck. Automating that (rather than hand-tuning prompts) is a practical lever for making agents more reliable.