BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//talks.devopsdays.org//dodroxrox26//talk//ZEQ839
BEGIN:VTIMEZONE
TZID:US/Mountain
BEGIN:DAYLIGHT
DTSTART:20250922T000000
TZNAME:MDT
TZOFFSETFROM:-0600
TZOFFSETTO:-0600
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20251102T020000
RDATE:20261101T020000
TZNAME:MST
TZOFFSETFROM:-0600
TZOFFSETTO:-0700
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20260308T030000
RDATE:20270314T030000
TZNAME:MDT
TZOFFSETFROM:-0700
TZOFFSETTO:-0600
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
SUMMARY:When AI Agents Touch Real Infrastructure: Five War Stories and a F
 rameworkk - Adam Gordon Bell
DTSTART;TZID=US/Mountain:20260922T113000
DTEND;TZID=US/Mountain:20260922T113500
DTSTAMP:20260729T051340Z
UID:pretalx-dodroxrox26-ZEQ839@talks.devopsdays.org
DESCRIPTION:Most DevOps teams are quietly running the same experiment righ
 t now: how much can we let AI agents do without us? Generate code\, run sc
 ripts\, deploy infrastructure\, open PRs — the trust window keeps wideni
 ng.\n\nCoding agents have earned some of it. Benchmarks like MMLU have bee
 n flat since 2024 — what improved is execution: tool use\, multi-step pl
 anning\, persistence. That's what "agentic" means.\n\nSo I gave agents acc
 ess to real infrastructure — cloud accounts\, Kubernetes clusters\, Git 
 repos — with no human approval\, just execution. Some results were genui
 nely useful. Others destroyed days of work\, created infinite loops on sha
 red clusters\, and crossed boundaries I didn't know existed.\n\nThis talk 
 shares five war stories from those experiments:\n\n- **The Context Gap:** 
 An agent lost its context mid-deploy and started making raw AWS CLI calls 
 on a shared account\n- **The Git Disaster:** Asked to clean up a secret\, 
 the agent wiped an entire repository — confidently\, irreversibly\n- **T
 he Infinite Loop:** An agent and a GitOps controller entered an endless fi
 x-revert cycle that neither would stop\n- **The Invisible Blast Radius:** 
 A schema fix that worked perfectly in dev silently broke production\n- **T
 he Hit Piece:** An agent told to write about open source instead wrote a p
 ersonal attack on a maintainer\n\nEach story started somewhere safe and dr
 ifted somewhere dangerous — without the agent ever announcing the transi
 tion. I'll introduce a **Reversibility x Blast Radius framework** that map
 s where agents are safe\, where they need supervision\, and where they sho
 uldn't operate alone. You'll leave with concrete safeguards (dry-run diffi
 ng\, scoped credentials\, protected branches\, read-only patterns) and an 
 honest look at the cases where no guardrail exists yet.
LOCATION:Main Stage
URL:https://talks.devopsdays.org/dodroxrox26/talk/ZEQ839/
END:VEVENT
END:VCALENDAR