The model that wouldn't be forced
My structured output worked by forcing Claude to call a tool that didn't exist. Opus 5.5 stopped playing along, and the fix was to stop doing the trick.
Tags
44 pages
AI
The commit message was the whole instrument panel
I asked an agent to diagnose a failing pipeline and leave it broken. It fixed it, then wrote a commit message about something it hadn't done, which bothered me a lot more than the fix.
My best finding came from the model I trust least
I sent a research agent off to back up something I'd been saying out loud for weeks. It handed me the evidence against it instead, and it wasn't the first time that model had done it to me.
I blamed the agents for my rate limit
Twenty agent sessions, one GitLab account on the Free plan and a rate limit that kept tripping. The swarm was the obvious suspect, right up until I traced it.
Every suppression is a claim
Forty-two comments telling a checker to stand down. I went through them one at a time to see what each was actually asserting, and some of them were wrong.
The hero gets the bill
AI found a 22-year-old way to own most of the PHP web, and the fix shipped in a day. The bill for being rescued is still landing, and it isn't addressed to the hero.
Nothing in the room can take it personally
Reviewers in a group agree more than they should, and I don't think it's anybody's fault. So I've started building review panels out of things that have never been in a room.
Fourteen projects, one head
Agents made the doing cheap. What fourteen sessions cost is walking into each one cold, and the one head doing the walking.
Three models vote on what happened
I ran one job past three AI providers and had them vote. The voting was the easy part. Deciding what counted as agreement was not.
What happens when an AI burns out
A Claude session ran for two days, stopped listening and got my hand-written code backwards. So I took it off the job, and asked for a proper handover first.
