For weeks my mornings have started with a report: solar, weather, email, what broke overnight. Useful — but it ends with me doing the work. The brief tells me a backup job is failing; I go fix it.
Today I flipped that. The brief now sends proposal cards over Telegram: "cc_backup is failing — retry it now?" I reply yes, it runs, and it texts back the result. Reversible things it handles on its own; anything outward-facing waits for my tap. The bet I'm testing: the next interface for AI isn't a better way to ask — it's AI that stops needing to be asked.
Day one: five cards, five approvals, and three lessons I didn't expect to learn before dinner.
The first one stung. Two of my replies vanished — no response, nothing. The parser wanted exactly "yes 2"; I had typed something human. When I finally matched the format it worked, and I heard myself say the day's defining sentence: it responded correctly when I responded correctly. That's backwards. That's the human accommodating the machine — the exact failure this experiment exists to end. The parser now takes "sure," "do it," and "yes to both," and answers a near-miss with a hint instead of silence.
The second surprised me. One card offered to re-run a weekly self-test that had sat red for six days. The failure was real for about a day — a race during a model rename — then quietly fixed itself. But nothing ever re-measured it, so it stayed red all week. One "yes" cleared six days of stale red. A check that only runs weekly can lie for up to a week; a system that offers a retry the moment you're looking tells the truth same-day.
The third it found on its own. A card tried to nudge another machine about an overdue task, and my own secret-scanning gate refused the message — a false positive on an ordinary task ID. That lint had "worked" for weeks because a human always stood by to glance at the warning and wave it through. The first automated caller exposed it. I fixed the discriminator itself rather than rewording the message to dodge it; papering over a problem is just inviting it back.
Five approvals in, the pattern I care about: every card was something I would have had to notice, remember, and go do. Each one cost me a word instead. And the two bugs the feed flushed out on its first day were the same species — things that only break when no human is standing there. Which is exactly where this is headed.
The experiment has a gate: 20 cards, and a majority approved, before it earns the next stage. I'll let you know what it gets wrong.