AWS details AgentCore optimizer using agentic reflectors for prompt tuning

AWS evaluated two reflector architectures: a Single Agent Reflector and an experimental Sub-Agent Reflector. The Single Agent Reflector prioritizes efficiency, achieving an 81.55% success rate on the AppWorld benchmark in just six minutes, significantly faster than baselines like GEPA and MIPROv2. In contrast, the Sub-Agent Reflector, which uses a swarm of agents to analyze traces independently, achieved the highest quality score of 95.83% on AppWorld. This multi-agent approach is better suited for complex scenarios with diverse failure modes, though it requires more computational resources.

To prevent overfitting and safety drift, the system applies strict guardrails, including length ceilings and checks for verbatim trace reuse. The Sub-Agent Reflector is available as an open-source experimental feature in the Strands GitHub repository, while the Single Agent Reflector is part of the managed AgentCore service. AWS recommends starting with small trace sets and using multi-modal reward signals to enhance the optimization process.