Amazon Bedrock AgentCore adds a feature for automatic optimization of agent system prompts
Amazon Bedrock AgentCore now offers optimization of agent system prompts: it uses production traces to suggest a revised prompt with an explanation, which is validated through offline evaluation and A/B testing before deployment.
Amazon released a feature for Bedrock AgentCore that automatically optimizes agent system prompts. The tool analyzes production traces of agent behavior recorded through AgentCore Observability together with an evaluation signal and uses them to suggest a revised version of the system prompt, accompanied by an explanation of why the change should improve agent quality. According to the company, the proposed changes need to be reviewed and tested through offline batch evaluation and online A/B testing on live traffic before they can be deployed.
At the core of the system is a so-called agentic reflector, which compares successful and unsuccessful agent runs and looks for patterns that determine the difference between them. The company describes two variants: Single Agent Reflector, which currently powers recommendations in AgentCore optimization and reviews the entire set of traces in a single pass, and the experimental, open source Sub-Agent Reflector, released as a preview in the Strands repository on GitHub. The latter uses a group of sub-agents, each of which analyzes one trace in its own context, with an orchestrator then consolidating the results into a single set of configuration changes.
To prevent unwanted side effects of optimization – according to the company, prompts tend to grow longer after optimization, incorporate quotations from specific traces, or weaken safety constraints to achieve a higher evaluation score – each proposed candidate passes through a set of rule-based guardrails before acceptance.
The company states that Single Agent Reflector and Sub-Agent Reflector were compared with the existing methods GEPA and MIPROv2 on two publicly available benchmarks. The source text does not include specific results from this comparison. You can find details in the source article.
Why it matters
Manually tuning agent system prompts requires reviewing long operational traces and repeated reassessment – the tool automates this process by identifying patterns of success and failure and suggesting a specific change along with a rationale. With guardrails and mandatory A/B testing before deployment, the decision to accept the change remains with the team, reducing the risk of a prompt with weakened safety constraints reaching production solely because of a higher evaluation score.
Two audiences, two different impacts
What this means
For individuals
Developers building agents on Amazon Bedrock AgentCore no longer need to manually review long operational traces and tune prompts component by component – the tool suggests a specific change to the system prompt, along with a rationale, that can be reviewed before deployment.
For a business
Companies running agents on the Amazon Bedrock AgentCore platform can reduce the cost of manual prompt tuning because optimization automatically suggests changes based on production data and validates them through offline evaluation and online A/B testing, shortening the agent improvement cycle.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.