Skip to content

AgentCore Integration

AgentBridge treats Amazon Bedrock AgentCore as a deployment and operations target, not as another portable orchestration framework. A LangChain or LangGraph AgentSpec can be deployed through the AgentCore CLI or infrastructure tooling, then invoked through the optional agentbridge-agentcore plugin. The plugin keeps Runtime invocation normalized while preserving native AgentCore services.

Coverage Contract

AgentCore surface AgentBridge connection Evidence
Runtime invocation AgentCore backend and AgentCoreClient.invoke_runtime() Offline adapter contract test
Memory memory_id binding and normalized metadata/payload Adapter test
Gateway and MCP gateway_url binding plus native operation escape hatch Config contract
Identity identity_provider binding Config contract
Observability trace ID forwarding and observability metadata Runtime client contract
A2A and AG-UI protocol labels preserved for the deployed runtime Capability metadata
Future AgentCore APIs AgentCoreClient.call(service, operation, ...) Dynamic native operation contract

Install the AWS client only when this integration is used:

pip install -e 'plugins/agentbridge-agentcore[aws]'

AgentCore deployment publishing, IAM, containers, CloudWatch destinations, and account-level governance remain AWS control-plane responsibilities. This boundary is deliberate: AgentBridge connects application code to those services without inventing a second deployment platform.

The credential matrix includes an opt-in agentcore_runtime lane. It requires AWS credentials available to boto3, AWS_REGION, and AGENTBRIDGE_AGENTCORE_RUNTIME_ARN; it is never run by the default test suite.

End-To-End Shape

flowchart LR
    spec["AgentSpec\nLangChain or LangGraph"] --> deploy["AgentCore CLI/IaC\nAWS deployment"]
    deploy --> runtime["AgentCore Runtime"]
    spec --> bridge["AgentCore plugin\nnormalized invocation"]
    bridge --> runtime
    runtime --> memory["Memory"]
    runtime --> gateway["Gateway / MCP"]
    runtime --> otel["Observability / CloudWatch"]
    runtime --> protocols["A2A / AG-UI"]
    langsmith["LangSmith"] -. tracing/evals/prompts .-> spec
    langfuse["Langfuse"] -. callbacks/API .-> spec

LangSmith and Langfuse

The LangChain plugin covers native tracing callbacks, prompt versioning, dataset/evaluation publishing, arbitrary LangSmith JSON/SSE APIs, Langfuse callbacks, and a dependency-free Langfuse JSON/SSE transport. AgentCore observability is an additional deployment-side signal path; it does not replace LangSmith or Langfuse, and all three can be enabled through the same AgentSpec.

Use examples/framework_to_agentcore_report.py to produce a no-network comparison artifact for local LangChain/LangGraph results and the configured AgentCore target.