Skip to content
BrPy

Amazon Bedrock Agents: Architecture Diagrams

Diagrams of AgentCore Runtime, the managed harness, Memory, Gateway and Identity, Strands multi-agent patterns, and Bedrock Agents Classic orchestration, action groups and sessions.

How the pieces fit together. Each diagram links to the page that covers it in detail.

One runtimeSessionId maps to one microVM. Reusing the ID returns to the same microVM and its in-process state until the idle timeout or maximum lifetime ends the session (AgentCore guide).

caller (app, other agent, MCP/A2A client)
│ InvokeAgentRuntime(arn, qualifier, runtimeSessionId, payload)
│ auth: SigV4 (IAM) or Bearer JWT (customJWTAuthorizer)
▼
┌──────────────────────────────────────────────────────────────┐
│ AgentCore Runtime │
│ endpoint "prod" ──► v4 endpoint DEFAULT ──► latest │
│ │
│ ┌─ session microVM (one per runtimeSessionId) ───────────┐ │
│ │ your code: BedrockAgentCoreApp on port 8080 │ │
│ │ GET /ping Healthy or HealthyBusy │ │
│ │ POST /invocations @app.entrypoint(payload, context)│ │
│ │ or an MCP server (/mcp) or A2A server (port 9000) │ │
│ └──────────────┬──────────────┬──────────────┬───────────┘ │
└──────────────────┼──────────────┼──────────────┼─────────────┘
▼ ▼ ▼
Bedrock models AgentCore Gateway, Code
(InvokeModel) Memory Interpreter, Browser
│
└── spans, logs, metrics ──► CloudWatch (Observability)
┌──────────────── AgentCore Runtime ─────────────────┐
InvokeHarness ───► │ managed harness │
(messages, │ harness.yaml: model, system prompt, tools, │
per-call │ skills, memory, limits, lifecycle hooks │
overrides) │ AWS runs the agent loop │
│ │
InvokeAgent ───► │ code-defined agent │
Runtime │ your container or zip: Strands, LangGraph, │
(payload) │ CrewAI, OpenAI Agents SDK... you own the loop │
└────────────────────────────────────────────────────┘
agentcore add harness ──► managed harness (harness.yaml)
agentcore add runtime / --template ──► code-defined runtime
agentcore export harness ──► harness converted to Strands code
agentcore add runtime --agent-id ──► Agents Classic agent converted to code

See Managed harness.

Short-term memory is the raw event log. Long-term records are extracted from it asynchronously by the strategies you enable, and stored under namespaces scoped by actor (and optionally session).

agent turn
│ add_turns(...) (CreateEvent: memory_id, actor_id, session_id, messages)
▼
┌──────────────────────────┐ async extraction ┌─────────────────────────────┐
│ short-term: events │ ───────────────────► │ long-term: memory records │
│ per actor + session │ strategies: │ /users/{actorId}/facts │
│ expire after N days │ SEMANTIC │ /users/{actorId}/... │
└──────────────────────────┘ USER_PREFERENCE │ /summaries/{actorId}/{sid} │
▲ list_events SUMMARIZATION └──────────────┬──────────────┘
│ (recent turns) EPISODIC │
│ search_long_term_memories(namespace_path, query)
└──────────── agent context for next turn ◄───────────────────┘

With Strands, AgentCoreMemorySessionManager makes these calls for you (Strands guide).

inbound auth (JWT or IAM)
agent ──── MCP over HTTP ────► ┌──────────────────────────┐
(MCPClient(url=gatewayUrl)) │ AgentCore Gateway │
│ tools/list, tools/call │
│ semantic tool search │
│ optional Cedar policies │
└───┬─────────┬─────────┬──┘
outbound creds: │IAM role │ OAuth / │ API key
▼ ▼ ▼
Lambda fn OpenAPI / MCP server,
Smithy API knowledge base
AgentCore Identity
inbound: user JWT ──► Runtime/Gateway authorizer ──► workload access token
outbound: workload token ──► token vault ──► OAuth token or API key for a tool
(@requires_access_token, @requires_api_key)

See Integration patterns and Identity.

agentcore/agentcore.json ─┐
agentcore/aws-targets.json├─► agentcore deploy --target prod
app/<Agent>/ (code) ┘ │
▼
CDK app (agentcore/cdk) ──► CloudFormation stack
│
┌─────────────┬───────────┼────────────┬──────────────┐
▼ ▼ ▼ ▼ ▼
IAM execution Runtime + Memory Gateway + Online eval
role endpoints targets configs
(CodeZip bundle, or a container image in ECR)

See the production guide.

agents as tools swarm graph (GraphBuilder)
─────────────── ───── ────────────────────
supervisor ┌──────────┐ handoff ┌───────┐ collect ─► validate ─┬─► analyse ─► report
│ │ │researcher│────────►│analyst│ ▲ │
▼ ▼ └──────────┘ └───┬───┘ └── FAIL (bounded)┘
sales billing ▲ handoff │
.as_tool() .as_tool() └───────────── reviewer ──► answer
model decides one agent active at a time; edges with conditions; ready
which to call shared context, handoff tool nodes run in parallel

See the Strands guide and multi-agent patterns.

These cover Agents Classic, which has been in maintenance mode since 30 July 2026.

invoke_agent(agentId, agentAliasId, sessionId, inputText, sessionState)
│
▼
┌─ Bedrock Agents (managed) ─────────────────────────────────────────┐
│ pre-processing (optional) ─► orchestration loop ─► post-processing │
│ │ model reasons, picks an action │
│ ┌────────────────┼──────────────────┐ │
│ ▼ ▼ ▼ │
│ action group: action group: knowledge base │
│ Lambda executor RETURN_CONTROL Retrieve (vector store) │
│ │ │ │
│ ▼ │ guardrail on input and output │
│ your Lambda fn │ │
│ (messageVersion 1.0 │ │
│ response envelope) │ │
└──────────────────────────────┼─────────────────────────────────────┘
▼
completion stream: returnControl event
your app runs the call, then invoke_agent again with
sessionState.returnControlInvocationResults

Agents Classic keeps the conversation inside the service. You don’t provide a database for it.

sessionId ──► session context held by Bedrock Agents
kept until idleSessionTTLInSeconds passes with no calls
sessionAttributes (passed to action-group Lambdas)
promptSessionAttributes (added to the prompt)
memoryConfiguration (optional, SESSION_SUMMARY)
at session end ──► summary stored per memoryId, kept storageDays
next session with the same memoryId ──► summaries added to context

See the Classic reference.

┌─────────────────────────────┐
user ─────────► │ supervisor agent │
│ agentCollaboration: │
│ SUPERVISOR (plans) │
│ SUPERVISOR_ROUTER (routes)│
└───┬─────────────┬───────────┘
│ associate_agent_collaborator (alias ARN, instruction)
▼ ▼
collaborator collaborator
agent alias agent alias
(own action groups and knowledge bases)

See multi-agent collaboration.