Multi-Agent AI Systems & the A2A Protocol

Hands-on workshop — Revision 1.2 — 08/01/26

This repository contains the labs, code, and setup for the Multi-Agent AI Systems & the A2A Protocol workshop. In it, you'll build multi-agent systems with LangGraph and CrewAI, learn Google's A2A (Agent2Agent) protocol v1.0, and connect agents built in different frameworks into a single cross-framework agent network.

Prerequisite: Experience building single AI agents (via the AI Accelerator, ai-aip workshop, or equivalent). Comfort with Python and LLM APIs.

---

Setup — GitHub Codespaces (recommended)

1. Click the button below, or use the green Code button → Codespaces tab → Create codespace on main.

   

2. Wait for setup to finish. The codespace installs Python dependencies, installs Ollama, and pulls the llama3.2:3b model. This takes 3-5 minutes after the codespace opens. When the terminal shows Ollama is ready with model llama3.2:3b, you're good to go.

3. Set up a free Groq API key (recommended — see below).

4. Open labs.md and start with Lab 1.

System requirements (Codespace machine type)

• 4 cores / 16 GB RAM / 32 GB storage (the default machine type configured for this repo)

Recommended: a free Groq API key

Every lab runs two ways, and the lab steps are identical either way:

	Setup	Speed	Answer quality
Groq (recommended)	free API key	seconds	reliable
Local Ollama (fallback)	nothing — works out of the box	30s to 2+ minutes	small model; often approximate

The labs default to the local llama3.2:3b model so they work with no account and no key. That model is only 3 billion parameters, though, and it shows: it will sometimes pass a sloppy argument to a tool, or ignore what a tool returned and answer from imagination. The lab still works — you'll see the agents delegate and the protocol do its job — but the answers are often wrong.

With a Groq key the same labs run in seconds and return correct answers, which makes the multi-agent behavior much easier to follow. Grab a free key (no credit card) at https://console.groq.com and set it in each terminal you use:

export GROQ_API_KEY=<your key>

Every lab file checks for that variable and switches automatically — there is nothing else to change.

Instructors: add GROQ_API_KEY as a Codespaces secret (github.com → Settings → Codespaces → Secrets, with access granted to this repo) and it will be present in every student codespace with no per-terminal export.

Alternative: local dev container

If you prefer to run locally in VS Code with Docker installed, use Dev Containers: Clone Repository in Container Volume... from the command palette and point it at this repo. The same setup scripts run.

Troubleshooting

• ollama: command not found — the setup script didn't finish. Run bash scripts/startup_ollama.sh and watch for errors; it needs zstd installed (the script installs it for you).
• Model responses hang — first query after startup loads the model into memory (1-3 minutes). Subsequent queries are faster.
• address already in use when starting an A2A server — a server from a previous lab is still running. Find and stop it: kill $(lsof -t -i:9999) (substitute the port from the error message).
• Package errors — re-run pip install -r requirements.txt inside the activated environment.

License

Materials in this repository are for educational use only by attendees of our workshops.

(c) 2026 Tech Skills Transformations and Brent C. Laster. All rights reserved.