Amazon Bedrock AgentCore now includes a system-prompt optimizer that examines scored production traces and proposes configuration changes for AI agents. The workflow connects AgentCore Observability with evaluations, offline batch tests and live A/B experiments so teams can review a recommendation before promoting it.
The current Single Agent Reflector searches a directory of traces, compares successful and failed runs and returns a coherent set of prompt edits. An experimental open-source Sub-Agent Reflector assigns individual traces to separate agents, then combines their diagnoses. Amazon says the single-agent version reached 81.55% on AppWorld in six minutes and 78.31% on WebShop in one minute. The multi-agent version reached 95.83% and 79.15%, respectively, but uses more computation.
Candidate prompts pass three guardrails: growth beyond 20% is rejected, safety constraints are checked, and exact phrases from traces cannot be copied into the new prompt. Those controls are intended to limit prompt bloat, weakened safeguards and superficial overfitting. Benchmark gains do not guarantee improvement on a company’s own workload, and live experiments still need careful monitoring. The useful change is procedural: prompt tuning becomes a testable recommendation pipeline with human review rather than an engineer manually reading long traces and editing instructions by instinct.