Agents in CI: How Autonomous AI Cleans Up 3,000 Flags Nobody Wants to Touch

Every frontend monorepo has a feature-flag graveyard. Ours had 3,212 flags. About 900 of them were fully rolled out ages ago and just sitting there doing nothing, and plenty more had dependencies tangled up in ways nobody remembered. The result was slower builds, worse startup, and every now and then an actual production incident. And nobody ever cleans these up by hand. Be honest, you won't either.

Outreach sells AI agents that do the work for sales teams. So I built the same thing for us engineers: agents that run in CI, hunt down the dead flags, check that it's actually safe to delete them, open the cleanup PR, file the Jira ticket, and ping Slack. Not "here's what AI thinks you should do" — the agent does it and hands you a PR to merge.

I'll show you how it's wired together (GitHub Actions plus an AI agent, talking to six different systems), why you can't just write a script for this (checking that a delete is safe turns out to be the whole game), and how something I built for one job ended up as a platform other teams started hanging their own automations off of.

And I'll be straight about the part that surprised me most: the hardest problem wasn't technical at all. It was getting people to actually care. We'll talk about that too.

No hype. Just patterns you can take home and build your own CI agents with.