Amazon Bedrock AgentCore

Definition

Amazon Bedrock AgentCore is a set of managed services for deploying production AI agents at scale. Announced at re:Invent 2024 and reaching GA in 2025, AgentCore decouples agent infrastructure from the agent logic — you can use any framework (LangGraph, CrewAI, Strands, LlamaIndex, custom) on top of AgentCore primitives.

AgentCore primitives

AgentCore Runtime — Serverless, session-isolated execution for agents. Handles concurrency, isolation per user session, and scaling.

AgentCore Memory — Short-term (session) and long-term (cross-session) memory with summarization, semantic recall, and privacy controls. Eliminates the need to build custom vector-store-plus-summarization plumbing.

AgentCore Gateway — Wraps existing APIs, AWS services, and Lambda functions as agent-callable tools with auth, throttling, and audit logging.

AgentCore Browser — Headless browser tool for web research and form-filling tasks, with anti-bot fingerprint management.

AgentCore Code Interpreter — Sandboxed Python/JS execution for data analysis and code generation tasks.

AgentCore Identity — Agent-to-service authentication using IAM Roles Anywhere and OAuth, so agents call APIs as themselves with auditable identities.

AgentCore Observability — Trace every agent step (model call, tool invocation, memory write) into CloudWatch and OpenTelemetry. Spotlight for failures and replays.

When to use AgentCore

• You have an agent prototype working locally and need production infrastructure.
• You want to keep your existing framework (LangGraph, CrewAI, Strands) but offload memory and tool plumbing.
• You need session isolation, identity, and audit trails for compliance (HIPAA, SOC 2).
• You want managed observability and replays of agent runs.

When not to use AgentCore

• Single-turn LLM call — use the Bedrock InvokeModel API directly.
• Knowledge-base lookup only — Bedrock Knowledge Bases is the lower-friction primitive.
• Sub-second latency requirements — AgentCore Runtime cold start is in the seconds range.

Common mistakes

Mistake 1: Skipping AgentCore Memory and building bespoke conversation memory in Lambda + DynamoDB. AgentCore Memory handles summarization, recall ranking, and PII redaction — building this yourself is months of work.

Mistake 2: Granting AgentCore Identity broad IAM roles. Each agent action should be scoped to the minimum permissions the underlying tool needs. Use AgentCore Gateway to enforce per-tool authorization.

Mistake 3: Treating AgentCore as a model — it is a runtime layer. You still pick the underlying foundation model (Claude, Nova, Llama) for reasoning.

Related AWS Services

• Amazon Bedrock — Foundation models for agent reasoning
• AWS Lambda — Backend logic invoked through AgentCore Gateway
• Amazon CloudWatch — Logs and metrics from AgentCore Observability
• IAM Roles Anywhere — Identity backbone for AgentCore Identity

Related FactualMinds Content

• Amazon Bedrock Consulting
• Generative AI on AWS