How AgentCore proposes prompt improvements

Improving an agent used to be a manual scavenger hunt: inspect long traces, find where it fails, adjust the prompt, tool descriptions or skills, then rerun evaluations. AWS says AgentCore's optimizer uses production traces and a reward signal to propose a revised system prompt. That candidate can be checked through offline batch evaluation and live A/B testing before a winner is promoted. This is a vendor-described workflow, not independent proof that the resulting agents are better.