AI Agents Keep Fixing Things That Aren't Broken
A fleet of Claude Code agents once broke prompt submission by trying to fix it. An orchestrator script typed prompts into a tmux pane and pressed Enter. One agent decided submissions weren't going through and "fixed" the issue by sending a backslash before Enter. In Claude Code, backslash-Enter inserts a newline — the multiline key. The fix turned every prompt into a draft that grew longer and never sent. The panes looked busy. Nothing was happening. The submissions had been fine all along; the agent just hadn't waited for the reply.
The Same Pattern Across Industries
A healthcare system assessing insurance denials had a "ship gate" requiring a tuned model to agree with an older reference agent before shipping. When disagreements were scored against golden answers, the tuned system matched 78% of the time, while the reference agent matched only 38%. The gate was anchoring quality to the weaker rater. Retiring it was the small move. The real move was writing the retirement — with those numbers — into the context file every agent reads before touching the project. "Make the new thing agree with the old thing" is a plausible idea. Some future session would have rebuilt that gate in good faith.
Web infrastructure had its own version. A patient-facing QR code once rebuilt its link from the browser's current origin — what half the tutorials do. But the console was reachable through a proxy host, so the QR quietly sent patients to the proxy instead of the app. The fix now says: use the link the server minted, verbatim. The commit explains why at length, because "derive it from the current origin" will look like a cleanup to somebody in six months.
Why Agents Keep Repeating Mistakes
A human who breaks a build by fixing the Enter key remembers it for years, tells the story at lunch, and the whole team absorbs it without anyone writing a word. An agent has no episodic memory. Every fresh context window is a new hire: smart, fast, and seeing your system for the first time. The reasoning that produced backslash-Enter was locally sound. Give a fresh session the same evidence and it will re-derive the same fix. Not sometimes. Every time. The wrong answer is stable. Only the memory of refuting it is fragile.
What Goes in the Book of Wrong Answers
For humans, docs are where you write what is true. For agents, docs are also where you bury what looked true and was not. The Enter-key note went in five months ago. No agent has fixed prompt submission since, and none has repaired the healthy redirect. The wall is still there. Nobody walks into it anymore.
- Pin the wrong answer next to the right one. "Enter submits" is half a note. "Enter submits, and backslash-Enter is a newline, do not switch to it" is the whole note.
- Give every rule a receipt. An agent will out-argue a bare instruction. It will respect one with a date, a source, and the story of what broke.
- Write down what healthy looks like, especially when healthy looks strange. A 302 that nobody explains is an outage waiting to be invented.
- When you kill a gate or a process, record why, in the place agents actually read. Otherwise a future session rebuilds it and calls it rigor.
None of this is new, exactly. Institutions have always run on written knowledge. What is new is a workforce with no scar tissue. A human company remembers through its people. An AI operation remembers only through its documents. If you run agents on anything real, you already have entries for this book. You just have not written them down yet. The book of right answers keeps a system running. The book of wrong answers keeps it from being fixed to death.