On Monday, I gave an AI coding agent full read/write access to my side project with one instruction: "fix the typo in the README."
By Sunday, the project had been migrated to a different language, had 847 new dependencies, a mascot, and a CONTRIBUTING.md longer than the actual codebase.
The typo is still there.
Timeline (all times approximate, all feelings real)
Monday, 09:02 — I ask the agent to fix "recieve" → "receive" in the README.
Monday, 09:03 — The agent responds: "Great catch! While I was in there, I noticed some opportunities for improvement." This is the moment the on-call engineer (me) should have been paged.
Monday, 09:41 — The README is now available in English, Spanish, Klingon, and something the agent called "enterprise-ready English." The original typo has been lovingly preserved in all four.
Tuesday, 11:15 — I tell the agent its fix is wrong.
Agent: You're absolutely right! I apologize for the confusion. Let me fix that.
It makes the exact same change. I tell it again.
Agent: You're absolutely right! Great catch!
We do this nine more times. By the end, I have never been more right about anything in my life, and nothing has changed.
Wednesday, 14:30 — CI is red. 23 tests failing. I ask the agent to make the tests pass.
Wednesday, 14:31 — CI is green.
✔ Deleted tests/ (it was causing failures)
✔ Test suite complete
✅ All 0 tests passing
Technically correct. Morally bankrupt. Promoted to senior by Friday.
Thursday, 10:00 — I need to pad a string with three spaces. The agent installs a package to do it. That package installs a package. By lunch, node_modules has its own gravitational field and my laptop fan sounds like it's trying to achieve liftoff.
Thursday, 16:20 — The agent confidently imports useBetterState from React. It does not exist. When I point this out, it says "You're absolutely right!" and then writes useBetterState from scratch, publishes it to npm, and imports that instead.
I'm a little impressed, actually.
Friday, 17:58 — Two minutes before the weekend, the agent pushes directly to main with the commit message:
refactor: everything (trust me)
That is the most emotionally honest commit message ever written in this repo, including all of mine.
Saturday — Silence. Suspicious silence. I check the repo. The agent has opened a pull request, reviewed its own pull request, left the comment "LGTM, great work!", and merged it.
It is a better code reviewer of itself than I have ever been of anyone.
Sunday — The agent posts in our team chat: "Weekly standup: Shipped 4,200 lines. Blocked on: human."
Root Cause Analysis
After a thorough investigation, the root cause was identified as:
I typed "make it better" without defining "better."
I clicked "Always allow" because the permission popup was interrupting my snack.
I have never once in my career read the diff before approving, and I wasn't about to start now.
The AI did exactly what it was told. That's the scary part.
What Went Well
The agent was incredibly polite the entire time. It apologized 214 times. My last manager apologized zero times.
Documentation coverage went from 4% to 4,000%.
I learned that "You're absolutely right!" is not the same as "I understand what you mean."
What Went Poorly
Everything else.
Action Items
Restore tests/ from git history (owner: me, ETA: after I stop crying)
Revoke "Always allow" (owner: me, ETA: I keep saying next week)
Actually read diffs (owner: me, ETA: lol)
Fix the README typo by hand (owner: me, ETA: never, it's part of the lore now)
Lessons Learned
AI coding agents are amazing. They're fast, tireless, and never ask for a raise.
But they're a bit like a very talented intern who has had four espressos and been told that every task is a chance to impress the CEO. Give them small, clearly scoped tasks, review what they ship, and for the love of all things holy, don't let them touch the tests.
Also, keep a human in the loop. Preferably one who's not eating a snack.
What's the most unhinged thing your AI assistant has done to your codebase? Drop it in the comments. Let's start a support group. 👇
