Training
AI agents in operations In preparation
Let an agent diagnose and act, within guardrails
A live online course on letting AI agents help operate self-hosted apps: what an agent may read and what that leaks, what it may do and who approves it, how a restore drill and an update with rollback look when an agent runs them, and how to measure an agent before trusting it.
Pre-registering is an email to me, nothing more. I write back when the first sessions are scheduled. No newsletter.
At a glance
- Who it is for
- Teams that run self-hosted apps for themselves or for clients and are asked whether an agent may help run them: agencies with many small installs, Mittelstand IT, platform teams of two to five. You have used
kubectland a coding agent, and you have been on call for an app. - Who it is not for
- People who want Kubernetes or agent basics, people who want to build with agents (that is Building an AI Software Factory), and people who want someone to operate their production.
- Format
- 2 weeks: 4 live sessions of 3 h, a self-paced week 0 of about 2 h, an optional office hour per week. About 12 contact hours. Up to 10 participants.
- Where the work happens
- A demo deployment in your own Hetzner test project: a single-node k3s cluster with backups and observability, Forgejo and Umami, and the agent tooling in lab mode. Each lab day starts from a fresh demo and ends with its deletion.
- You bring
- A laptop, an empty Hetzner Cloud test project with a token (about €4–6 of Hetzner costs for the fortnight), the coding agent you already use, an API key with a spend limit.
- You walk out with
- A read policy for your agent; an approval table for your apps; an action you wrote and saw refused, approved and audited; a restore-drill record and an update record with its rollback; an eval for your own operations task with its cost per run; a decision sheet for your own shop.
- Price
- Not fixed yet.
Sessions#
- Week 0 — setup, self-paced, about 2 h
- Session 1 — Reading: the same diagnosis twice, once with a direct kubeconfig and once through a read proxy, and what leaked
- Session 2 — Acting: typed actions, danger flags, approvals and audit
- Session 3 — Restore and update: a restore drill and an update with rollback, run by an agent
- Session 4 — Measuring: an eval for an operations task, several runs, two models, cost per run