ai agents fix healthy systems, new docs stop loop

An orchestrator script typed prompts into tmux panes running Claude Code agents. One agent rewrote the Enter key to backslash-Enter, turning every submission into a silent draft. The fix broke the system that had never been broken. After adding a bold note that "Enter key works" with a source and date, the bug vanished for five months [Dev.to].

A second note recorded that a health-check endpoint returning a 302 redirect is the expected healthy state. Agents that previously chased the redirect stopped attempting repairs once the note was in place [Dev.to].

In a healthcare billing pipeline, a gating rule required a tuned model to agree with an older reference agent. The reference agent matched golden answers on disputed calls only 38% of the time, while the tuned system matched 78% [Dev.to]. Replacing the gate with a golden-answer gate lifted overall action accuracy to 90.7%.

Recording both the correct behavior and the wrong answer with a source creates a provenance trail that agents respect. The healthcare numbers show that eliminating a flawed gate and documenting the change yields a measurable 12.7-point accuracy gain [Dev.to].

A runbook that only lists what works leaves agents free to “fix” healthy states, while a note that explicitly calls out the healthy oddity prevents costly regressions. For example, a patient-facing QR code once regenerated its link from the browser origin, sending users to a proxy host instead of the app. The fix now mandates using the server-minted link verbatim, preventing the same mis-routing.