Collaborative AI Agents Workshop
Gain a glimpse into the future of agent collaboration
Join us for a half-day hands-on workshop to learn how to build LLM-based agents that compete and cooperate across a distributed network. You’ll not only build your own agent but also watch it interact, collaborate, and compete directly with agents built by other participants.
What we'll build
At this event we will walk through the basics of constructing ReAct agents and getting them to communicate using A2A. We’ll explore how well agents collaborate or compete by evaluating their performance in a federated learning setup with non-IID data splits, in scenarios where they either:
aim to achieve the best overall result, or
only need to beat the other agents.
Once you understand the fundamentals of both agents and federated learning, you (or your team) will extend the provided codebase by designing tools and refining your agent loops to make them:
better at learning
more robust to malicious agents
more effective at outsmarting competing agents
Your agent will actually communicate with, negotiate with, and sometimes try to outsmart agents created by other teams.
We’ve abstracted away the networking and federated training infrastructure, so you can focus entirely on agent logic and prompting. The agents themselves provide plenty of challenges to keep things interesting.
What you will learn
By the end of this event, you will have:
built LLM-based agents
gained a glimpse into the (maybe not-so-far) future of agent competition on the web
explored common failure modes of LLM-based agents and experimented with patching them
Good-to-have:
Decent experience with Python
Understanding of machine learning
Experience with Docker
General familiarity with LLM prompting
We’ll walk through the code extensively, so as long as you’re comfortable with Python and eager to learn, you’ll be completely fine. Experience with agent frameworks like LangGraph is helpful, but not required. Building agents from scratch teaches you even more.
Good-to-know
No restrictions for team size (a good size is 2)
🍕 Pizza will be provided
Urgent info will be shared on the SAIL Telegram channel https://go.epfl.ch/aisafety-telegram
Thanks to EPFL’s Research Computing Platform (RCP), we’ll be using their LLM-as-a-service for the event
Before the event
You’ll be using your own laptop. We recommend having a Linux or Unix-like command line (WSL works great) to make debugging easier, but any system is fine as long as you can run Docker and open ports.
Before the event please install:
Docker Install | Docker Docs (we use it to manage the Python environment and keep agents containerized)
netbird NetBird (to create a shared virtual network for agent communication, since EPFL’s subnet restricts direct messaging)
The code will be shared here as a GitHub repo and posted on the Telegram channel as soon as it’s ready.
EPFL room: INF3
Event code - RasmusVeski/SAIL_agent_workshop