Pattern: Evaluator-Optimizer Dual-Agent Loop Now Productionized in Claude Managed Agents

Anthropic·medium signal

The evaluator-optimizer pattern — separating the 'doer' agent from the 'judge' — has moved from design-pattern theory to first-party infrastructure. Claude's Outcomes feature implements this with isolated context windows, rubric-based grading, and automatic re-iteration. For builders running their own agent loops, the key structural insight is that the judge must never share the doer's context, or it will rationalize the doer's mistakes instead of catching them.