Tip: Managed Agent Outcomes Uses an Isolated Grader Context Window — Decouple Your Judge From Your Doer

Anthropic·medium signal

Claude Managed Agents' Outcomes feature runs a separate grader in its own context window, completely isolated from the agent's reasoning chain. This prevents the grader from being influenced by the agent's justifications. Write rubrics describing success criteria; the grader evaluates, pinpoints gaps, and the agent iterates. Benchmarks show +8-10% improvement over single-agent approaches on document generation tasks.