Machine Learning
Amazon Bedrock AgentCore
Amazon Bedrock AgentCore is a platform for securely building and operating agents with managed runtime, memory, gateway, identity, policy, observability, evaluation, and built-in tool capabilities. It supports varied frameworks, models, MCP and A2A interfaces, and both single- and multi-agent workloads.
Key points
- Runtime provides isolated serverless execution, including extended asynchronous agent workloads
- Gateway converts Lambda, OpenAPI, Smithy, HTTP, agents, and model providers into governed agentic targets, including MCP-compatible tools
- Identity manages inbound and outbound agent authentication; Policy enforces Cedar authorization over Gateway tool calls
- Memory provides short- and long-term agent context but should remain separate from authoritative business records
- Observability and Evaluations use traces to analyze sessions, tool calls, task completion, and custom behavior
When to use it
- Operate a custom-framework agent with managed identity and memory
- Expose enterprise APIs as controlled MCP tools through AgentCore Gateway
- Evaluate multi-agent task completion and tool trajectories before and after release
Exam tips
- Select AgentCore for current production infrastructure around flexible agents; Bedrock Agents Classic remains a legacy managed-orchestration option for existing customers
- MCP is an interoperability protocol, not an authorization boundary; Gateway Policy and target-side validation still matter
- Memory, identity, and policy do not remove the need for retention rules, idempotency, human approval, and bounded execution