incident-multiagent

An AWS incident-response assistant: an operations-oriented chat that combines natural-language reasoning with CloudWatch access (via MCP tools) and web search (documentation and best practices).

This repository demonstrates a clear multi-agent pattern: an orchestrator delegates to specialists, deployed on Amazon Bedrock AgentCore and implemented with the Strands Agents SDK in Python.

---

Core concepts

Piece	Role
AgentCore	Managed service that runs the agent container, exposes the invocation endpoint, integrates memory, and can attach an MCP Gateway for external tools.
Strands	Python framework for agents (Bedrock models, tools, streaming). Every agent in this project is a Strands agent.
Orchestrator (supervisor)	The primary Strands agent: it does not do everything itself—it decides when to call the other agents as tools (agent-as-tool). That avoids a separate A2A protocol inside the same process.
Frontend	A React app that talks to the AgentCore runtime (streaming) and uses Cognito-aligned authentication from the deployment.

In practice: the user types in the chat → the AgentCore runtime runs main.py → the Strands supervisor interprets the request and, when needed, invokes the monitoring agent (MCP tools to CloudWatch) or the web search agent (Tavily).

---

Multi-agent architecture

flowchart TB
  subgraph users["Who uses the system"]
    U[User or on-call engineer]
    FE[React frontend]
  end

  subgraph agentcore["Amazon Bedrock AgentCore"]
    RT[Runtime: agent container]
    MEM[(Memory: short- and long-term)]
    GW[MCP Gateway → CloudWatch]
  end

  subgraph strands["Strands agents in one process"]
    SUP["Supervisor (orchestrator)"]
    MON["Monitoring agent"]
    WEB["WebSearch agent"]
  end

  U --> FE
  FE -->|"Streaming invoke + identity (session / actor)"| RT
  RT --> SUP

  SUP -->|"Delegation: agent as tool"| MON
  SUP -->|"Delegation: agent as tool"| WEB

  MON -->|"MCP tools"| GW
  WEB -->|"Web search"| TAV[Tavily API]

  SUP -.->|"Memory via AgentCore hooks"| MEM

Loading

How to read the diagram: everything inside the Strands block runs in a single container on one AgentCore runtime. The supervisor is the logical entry point for the user message; the other agents are specialists invoked by it.

---

Monorepo layout

Directory	Contents
backend/	AWS CDK (TypeScript), Docker image for the agent, Python Strands code and BedrockAgentCoreApp.
frontend/	Chat UI against the AgentCore endpoint.

---

Request lifecycle

1. The frontend sends the prompt to the AgentCore runtime ARN.
2. AgentCore starts or reuses the container and calls the entrypoint (@app.entrypoint).
3. The supervisor (Strands) reasons and, when appropriate, runs the tool that wraps the monitoring or web-search agent.
4. Events (text, tool use, errors) are streamed back to the client for the conversation and delegation UI.

---

Main technologies

• Amazon Bedrock (Claude models via Strands).
• Bedrock AgentCore (runtime, memory, MCP gateway as defined in the stack).
• Strands Agents SDK (agents, tools, streaming).
• AWS CDK v2 (aws-cdk-lib) for cloud resources.