Project Overview
DevOpsCrew is a multi-agent DevOps assistant where a supervisor delegates work to specialised sub-agents — source control, monitoring, code review, service health — with every write operation gated by a human.
Objective
Build a supervisor-worker graph in LangGraph with sub-agents exposed over MCP and discoverable via A2A, deployed to a managed agent runtime with policy-based access control.
Scope
- A supervisor-worker LangGraph with conditional routing and shared state.
- Sub-agents for source control, monitoring, code review, and service health.
- All sub-agents expose MCP servers (stdio for dev, SSE for production).
- A2A peer delegation between agents via standard endpoints.
- HITL interrupts before every write, with exact checkpoint resume.
Stack
- LangGraph for the supervisor graph and persistence.
- FastMCP-style MCP servers wrapping CLIs and cloud SDKs as tools.
- A2A discovery via well-known endpoints with JSON-RPC transport.
- SQLite state serialisation on interrupt.
- A managed agent runtime with microVM isolation, a tool gateway, short- and long-term memory, and per-agent identity.
Governance
- A policy language for agent permissions (writes require approval; read-only agents restricted from destructive actions).
- Tracing, metrics, and structured logging across all agents.
Demo Scenario
Triage a production incident → automated GitHub issue creation → code review by the code-review sub-agent → human approval → write performed — end-to-end in under a minute.
Deliverables
- LangGraph supervisor with sub-agents and shared state.
- All MCP servers implemented and registered with the gateway.
- HITL interrupt-and-resume working for all write operations.
- End-to-end A2A peer delegation.
- All agents deployed to a managed runtime with policies and observability live.
Prerequisites
Modules 20–23 (function calling/MCP, LangGraph, A2A and observability, AgentCore deployment).