Agents · Case study
Multi-agent orchestration with A2A
Specialist agents that discover each other and collaborate through Google’s A2A protocol — no hardcoded glue between Research, Writer, and a routing host.
Problem
You can build a Research agent and a Writer agent that are each excellent alone. Wiring them together usually means hand-rolled URLs, brittle glue, and spaghetti the moment a third specialist appears.
Approach
- Host-agent pattern: one Routing agent discovers specialists via Agent Cards at /.well-known/agent-card.json
- Specialists stay thin — LangGraph for reasoning, MCP for tools (DuckDuckGo, filesystem)
- A2A JSON-RPC for messaging so any language/runtime can join the mesh
- Session-scoped memory so multi-step “research then save” stays coherent
Outcome
A small orchestrator where adding a specialist is “deploy + register URL,” not rewiring every peer. Full walkthrough published on the blog with architecture diagrams and a working demo repo.
Stack
PythonLangGraphMCPA2A SDKGeminiFastAPI
More work