AWS adds AgentCore optimization tools for production AI agents

Amazon Web Services announced new optimization capabilities for Amazon Bedrock AgentCore that are designed to help teams improve AI agents running in production.

What AWS announced

In a June 2026 AWS announcement, Amazon Web Services said AgentCore now includes optimization features that use production traces to identify issues and recommend improvements for deployed AI agents. AWS said the capabilities are intended to support agents running in both AWS and non-AWS environments.

AWS described the feature set as part of a loop for observing, evaluating, and improving agent behavior. A separate AWS announcement from May 2026 said the AgentCore optimization preview combines recommendations with batch evaluations and A/B tests, giving developers a way to validate changes before or during rollout.

The AWS Machine Learning Blog similarly described the preview as a system that analyzes production traces, generates recommendations, and validates changes through batch evaluation and A/B testing.

How the optimization features work

According to AWS, the new AgentCore capabilities analyze production traces from deployed agents. Those traces can be used to surface insights about failures, user intent, and the trajectory an agent followed while attempting a task.

That matters because agent performance problems are often difficult to diagnose from a final answer alone. A trace can show which tools the agent used, what path it took, and where execution diverged from the expected result. AWS says AgentCore uses this information to generate recommendations for improving agent behavior.

AWS also highlighted batch evaluation as part of the release. Batch evaluation can be used to test agent changes against a defined set of tasks or examples, rather than relying only on ad hoc inspection. The company also said A/B testing is included, allowing teams to compare different agent versions in production-like or production settings.

Focus on deployed agents

The emphasis of the announcement is production operations rather than initial model selection or prompt creation alone. AWS framed the capabilities as a way to continuously improve agents after deployment, using real interaction data and validation steps.

The AWS sources also indicate that AgentCore is meant to work across AWS and non-AWS environments. That positioning is important for organizations that deploy agents across multiple systems and need performance monitoring and iteration mechanisms that are not limited to a single application stack.

Why it matters

As companies move AI agents from prototypes into real workflows, evaluation becomes more complicated. Agents may call tools, follow multi-step plans, interact with APIs, and handle varied user requests. A simple pass-or-fail check on the final response may not explain why an agent failed or whether a proposed fix improves reliability.

AWS is addressing that operational problem by tying together trace analysis, recommendations, batch evaluation, and A/B testing inside AgentCore. The company’s announcements do not claim that these tools eliminate the need for human review or domain-specific evaluation, but they do present them as a structured way to diagnose and test agent improvements.

Availability

AWS described the AgentCore optimization capabilities as being in preview in its May 2026 announcement and AWS Machine Learning Blog post. In its June 2026 announcement, AWS said AgentCore introduced new optimization capabilities for continuously improving agents in production.

The available AWS materials describe the capabilities at a product level: production trace analysis, failure, intent and trajectory insights, recommendations, batch evaluation, and A/B testing. They do not provide independent benchmark results in the cited excerpts, so performance claims should be evaluated in the context of a customer’s own agents, tasks, and deployment environment.