Ship production-grade agents on AWS.
7 modules · 37 lessons · 8 hours. Start at the top, or jump to what you need.
Your progress
Loading your progress…
Start the course
What Is AgentCore?
Foundations: Agents in Production
Start here. Understand the gap between an agent that demos and an agent that survives production, meet the seven AgentCore services and how they compose, see how AgentCore differs from the older Bedrock Agents, and set up your account, IAM, and the starter toolkit to deploy your very first agent.
What Is AgentCore?
The prototype-to-production gap, and the platform that closes it
12 minThe AgentCore Service Catalog
Seven GA services plus the preview frontier
13 minAgentCore vs. Bedrock Agents vs. Flows
Managed, modular, or visual — and when to pick which
12 minFramework- and Model-Agnostic by Design
Why a container that speaks a protocol changes everything
12 minPricing, Regions & the Two CLIs
Reading the consumption model and the fast-moving facts
13 minAgentCore Runtime
The heart of the platform. Package and deploy an agent with the Bedrock AgentCore SDK and the agentcore CLI, understand per-session isolation and long-running (up to 8-hour) execution, stream responses, host MCP and A2A protocols, and run the same agent whether it's built on Strands, LangGraph, CrewAI, or anything else — on Bedrock models or not.
The Runtime Object Model
Runtimes, versions, endpoints, sessions, and two planes
13 minSession Isolation & the Security Boundary
One microVM per session, and who owns the user mapping
13 minYour First Agent with BedrockAgentCoreApp
Four lines to a deployable HTTP service
14 minProtocols: HTTP, MCP, A2A & AG-UI
One container, four wire contracts
14 minStreaming, Long-Running & Async Work
SSE, the 8-hour ceiling, and not starving /ping
13 minLimits, Scaling & Runtime vs. Lambda
Quotas, the 2 vCPU/8 GB ceiling, and when to use what
12 minFrameworks, Models & Tools
An agent is your framework, your model, and the tools it can reach. Wrap any framework (Strands, LangGraph, CrewAI, OpenAI, ADK, LlamaIndex) for Runtime, run it on Bedrock or non-Bedrock models with Guardrails, then give it real capabilities: turn Lambda/OpenAPI/Smithy backends into secure MCP tools with AgentCore Gateway, surf the web with the managed Browser, and compute in the sandboxed Code Interpreter.
Wrapping Any Framework
The same agent across Strands, LangGraph, CrewAI, and more
14 minModels & Guardrails
Bedrock and beyond, with safety that spans providers
13 minAgentCore Gateway: The M×N Tool Problem
Turn APIs and Lambdas into MCP tools
13 minBuilding Gateway Targets
Lambda, OpenAPI, Smithy, MCP servers, and Runtime-as-tool
14 minGateway Auth & Semantic Tool Search
Ingress, egress, and finding the right tool among thousands
14 minThe Browser Tool
A managed, isolated cloud browser your agent can drive
13 minThe Code Interpreter
A sandboxed environment for code your agent writes
13 minAgentCore Memory
Stateless agents forget everything. Add short-term memory for coherent conversations and long-term memory that persists across sessions, choose the right extraction strategies (semantic, summarization, user preferences), and wire memory into your framework so your agent actually learns who it's talking to.
Memory Fundamentals
Short-term events, long-term records, and two planes
13 minMemory Strategies
Semantic, summary, user-preference, episodic — built-in to self-managed
14 minActors, Sessions, Events & Namespaces
The data model and namespace-scoped access
12 minUsing the Memory SDK
MemorySessionManager, events, retrieval, and the LLM turn helper
14 minMemory in Frameworks & at the Limits
Strands integration, the batching trap, and quotas
12 minIdentity, Security & Governance
Production agents act on real systems, so identity is non-negotiable. Master AgentCore Identity — inbound auth (who can invoke), outbound auth (the agent acting on a user's behalf via OAuth and the token vault), and provider integrations (Cognito, Entra, Okta, Google) — then govern behavior with the Cedar-based Policy engine, Bedrock Guardrails, least-privilege IAM, and defenses against agent-specific threats like prompt injection.
AgentCore Identity Overview
Agent identities, inbound vs. outbound, the workload directory
12 minInbound Auth
SigV4 vs. JWT/OIDC, and invoking a protected agent
13 minOutbound Auth & the Token Vault
The workload-access-token flow, 2LO vs. 3LO, refresh tokens
14 minCredential Providers & SDK Decorators
@requires_access_token, built-in vendors, and the callback gotcha
13 minGovernance: Policy, Guardrails & IAM
Cedar policy enforcement, and the Identity-vs-IAM split
13 minObservability & Operations
You can't operate what you can't see. Instrument agents with AgentCore Observability and OpenTelemetry, read traces, spans, and sessions in the CloudWatch GenAI dashboard, evaluate quality, then handle the operational realities — service quotas, cost, infrastructure-as-code (CDK/CloudFormation/Terraform), and CI/CD for agents.
The Observability Model
Metrics, spans, traces, logs, and sessions
12 minOTEL & ADOT: Auto vs. Manual
Runtime auto-instrumentation and instrumenting everything else
13 minSetup & Viewing Your Data
Transaction Search, the GenAI dashboard, and trace correlation
13 minEvaluations & Auditability
Scoring agent quality on the traces you already emit
12 minBuilding a Production Platform
Your capstone. Combine Runtime, Gateway, Memory, Identity, and Observability into a reference architecture for a real agentic application, choose single- vs multi-agent designs with A2A, engineer for reliability and scale, satisfy governance and compliance, and walk through a production go-live checklist end to end.
End-to-End Deploy with the AgentCore CLI
create, dev, deploy, invoke — CodeZip vs. Container
14 minProduction Hardening & IaC
VPC, KMS, least privilege, and infrastructure as code
14 minComposability & the Agent Platform
All-in or à la carte; building a paved path for teams
13 minThe Preview Frontier
Harness, Policy, Evaluations, Registry, Payments, Optimization
13 minCapstone: A Production-Grade Agent
Memory-backed, tool-using, observable, governed — end to end
15 min