Bedrock AgentCore now rewrites your agent's system prompt from production traces, hitting 95.83% on AppWorld

AWS Machine Learning Blog·medium signalXBlueskyLinkedInCopy link

AWS shipped optimization in Amazon Bedrock AgentCore, an agentic reflector that reads evaluated traces, contrasts successful against failed runs and proposes system prompt edits that must clear safety guardrails. The Single Agent Reflector reviews the whole trace corpus in one pass and reached 81.55% on AppWorld in 6 minutes and 78.31% on WebShop in 1 minute, reported as 18-36x faster than competing methods. An experimental open-source Sub-Agent Reflector analyzes traces independently before aggregating and reached 95.83% on AppWorld, a 23-point gain over baseline, and 79.15% on WebShop.