Amazon Bedrock: Custom Lambda-evaluators for AI Agent Control

Amazon Bedrock AgentCore now supports custom Lambda-based evaluators—functions that validate AI agent responses in real-time before they reach the user.

What Are Evaluators in AgentCore

Evaluators are validators that check the quality, safety, and factual correctness of responses generated by the agent. Amazon Bedrock already provides built-in checks for hallucinations and style, but the new functionality allows you to write custom checks tailored to a specific domain. Why? Because generic checks don't catch domain-specific errors. A financial agent needs to validate stock quotes for accuracy. A healthcare agent needs to ensure FDA guideline compliance. A legal agent needs to ensure compliance with local legislation.

How Lambda-evaluators Work

Custom evaluators live in AWS Lambda and integrate with AgentCore via API. When an agent generates a response, Bedrock sends it for validation before delivering it to the user. The Lambda function analyzes the text and returns a verdict: pass as-is, edit, or block.

Evaluators work in two modes:

• On-demand—validation on request, at DevOps' initiative
• Online—real-time validation integrated into the agent flow with no user-facing latency

The second mode is more useful in production: the system guarantees that bad responses don't get through, but users don't experience long validation delays.

Four Types of Checks

AWS demonstrated four Lambda functions for a financial market-intelligence agent:

• Fact-checking—comparing responses against verified sources using AWS services, such as DynamoDB with historical data
• PII detection—identifying personal data (SSNs, credit card numbers, phone numbers) using AWS Comprehend
• Real-time alerting—sending notifications via SNS, Slack, or webhooks when issues are detected
• Custom domain logic—validating the realism of financial figures, detecting contradictions, and enforcing business rules

Integration with AWS Ecosystem

The financial agent combines custom Lambda checks with Bedrock's built-in evaluators. Built-in checks catch hallucinations and grammar issues. Lambda functions catch finance-specific errors. Integration with other AWS services is native: Lambda can call AWS Comprehend, Kendra, SNS, SageMaker—all within a single ecosystem with no integration overhead.

What This Means

AWS is making a serious move toward human-in-the-loop AI for regulated industries. Previously, quality control was either manual or required proprietary solutions—now it's native in Bedrock. For the financial and healthcare sectors, this is critical: regulators require full transparency, auditing of all decisions, and the ability to block incorrect responses before release. Lambda-evaluators provide exactly that.