Hey Cloud, Deploy This: The Agent That Started Improving Itself (sponsored)
Michael Dudli
Serhii Tonkoshkuryk
15:15 - 15:30
La Cave
This year at Cloud Native Zürich we pitched a vision: infrastructure you talk to, Day 0 through Day 2. Since June, that agent has been running against a real cloud API. No mocks, no sandbox, no curated happy path,and somewhere in those three months it stopped being the thing we maintain and became the thing that maintains itself. We'll run it live: provision a Kubernetes cluster, have it explain the cluster back to us in plain English, audit it, then fix what it finds. Then the part worth your time: what broke, and what each break taught it. - Why every guardrail that actually held turned out to be an authorization boundary, not a prompt - The bug that only surfaced because something called our API a few thousand times without getting bored - Why "flag it" and "fix it" are separated by more engineering than any demo lets on The interesting part isn't that it works. It's that none of the improvement came from a roadmap meeting: it came from use. We'll show what compounds, what doesn't, and what it means now that we're shipping one sealed agent per customer, on Swiss infrastructure, with a permission set short enough to read out loud. For platform engineers, SREs, and anyone who's had to explain an agent's blast radius to a security team.