Agentic AI-Based Platforms - Part 3: Building Enterprise Operations with AgentCore Policy, Evaluation, and Observability

Amazon Bedrock AgentCore places Policy, Evaluation, and Observability at different points in time for enterprise operations. Policy blocks permissions before invocation, Evaluation checks quality after execution, and Observability collects traces, metrics, and logs across the entire span.

• Policy: Deterministically allows/denies Tool calls using Cedar and the Policy Engine.
• Evaluation: Verifies response quality with 13 built-in evaluators and LLM-as-a-Judge.
• Observability: Tracks execution flow with ADOT and the CloudWatch GenAI dashboard.

Policy defaults to Default Deny, with forbid taking precedence over permit. By combining principal, action, resource, and condition, rules such as department-based access restrictions, RBAC, and emergency lockdowns can be expressed at the Tool level. It is recommended to first check the policy's impact using LOG_ONLY, and switch to ENFORCE once no issues are found.

When JWT claims alone are insufficient, the Lambda Interceptor fills the gap. The Interceptor can process headers and bodies to inject external DB lookup results or decoded JWT information; use metadataConfiguration for simple pass-through, and passRequestHeaders when processing is needed. However, headers injected by the Interceptor have delivery constraints with Runtime and Lambda targets, and in case of header conflicts, the Interceptor's value takes precedence.