The people who wrote the code teach every course live, online. Each course has thirty seats and runs ten weeks. All three courses run in every cohort, and there are four cohorts a year. Your seat includes the recordings. The lab is a real cluster, and every student gets their own tenant on it.
You build GPU nodes, schedulers, model serving, and the boundary around the agents that call the models. We taught this course before. This run adds Slurm, the NVIDIA stack, and KServe.
Winter 2027 · starts 11 January 2027 Agents in ProductionYou learn how to run an agent fleet for real: identity, grants, sandboxed execution, budgets, and replay. The patterns come from Gibson.
Winter 2027 · starts 11 January 2027 Attacking and Defending AI AgentsThe first half breaks agents the way the published attacks do. The second half rebuilds one that refuses every one of them, and you measure it on AgentDojo before and after.
Each week has two live sessions of one hour and one lab. The first session builds the idea. The second picks up where the first stopped and runs the lab with you. Every lab is a numbered set of steps, so a student who missed the session can run it alone. The failures in the labs are on purpose. Half of the teaching happens when the cluster refuses a step and you learn why.
Everything you build lives in Git from week two. The shared environment is torn down after every session, and nothing in your namespace survives. What you carry out is a repository that brings it all back in one command.
There is no homework and no grade. Each week ends with an optional list of deeper reading and exercises. No later week depends on any of it.
Ten seats is one call. Send the team, and the last four labs run against your own stack.
A seat is also for sale through our partner Emage, at the same price.
We announce every run to the list before it goes public. Thirty seats go fast.