A2A Agent Discovery: Replace In-Process Calls with Agent-to-Agent Protocol

🛠 Lab from the AI Platform Engineering Bootcamp. Used in Week 4. Bootcamp landing page: https://academy.tekanaid.com/bootcamps/ai-platform-engineering-bootcamp Parent course(s):

• Week 4: AI Agents and Agentic Workflows (slug: ai-agents-104)

🟡 Beta bootcamp lab. Hands-on instructions, check scripts, and solve scripts are in place. Lab is part of the running Platform Assistant capstone arc that grows across all 8 weeks of the bootcamp.

In the previous multi-agent lab, your supervisor and specialist agents lived inside a single Python process and talked through in-memory function calls. That works for a demo but breaks the moment your platform team wants to deploy specialist agents independently, scale them on different hardware, or let other teams contribute their own agents to the catalog. Distributed orchestration needs a wire protocol, not a shared Python runtime.

This lab pivots the same multi-agent topology to the Agent-to-Agent (A2A) protocol. You will run two specialist agents (a fictional catalog agent and a fictional CI/CD agent) as independent HTTP services, each publishing an agent card at a well-known endpoint. Then you will build a supervisor that discovers those peers at runtime, reads their agent cards, picks the right specialist for an incoming prompt, and delegates the work over JSON-RPC. By the end you will have traced the full HTTP/JSON-RPC traffic between processes and understand exactly what changes when in-process orchestration becomes a network call.