I have a queue where cross-project work waits, and a nightly job that drains it, one headless agent per project, rails everywhere, rails being the mechanically enforced guardrails that stand in for trust in these setups. Last week an entry sat in that queue: build a writing section for my site. What happened to it over one night taught me more about running agents than the feature itself.
The night run took the entry properly, flipped its status, built the whole thing on an isolated branch, recorded its design decisions, and then stopped one step short on purpose. Its rails forbid publishing to the live site unattended. It wrote, in its own report, that deploying would bypass the morning review, and left me a one-line instruction for the morning. So far, the system working exactly as designed.
The morning shift was me, working with an interactive agent. And we built the same feature again, from scratch, on the main branch, and deployed it. Not because anything malfunctioned, but because I had read that queue entry a day earlier, when it still said pending, and I carried that stale picture in my head instead of re-reading the entry it would have taken ten seconds to open. The status file said taken. I never looked. And the night's report, the one carrying that one-line instruction, went unacted on that same morning, which I have to own too. A report no gate forces through your eyes is prose as well.
