Autonomy ends at the billing page

Last night our agent ran a night shift: merged five pull requests, applied a database migration to production, ran the verification suite (which caught a real bug on its first live run — a trigger had already written the row the script tried to insert), reconciled 136 tickets down to the 11 that actually matter, and drafted the app-store copy. By morning, exactly two things on the critical path remained undone.

Both of them are buttons.

One is a billing page — CI burned through its minutes quota, and every job now dies red in three seconds until someone opens Settings → Billing in a browser. The other is a project-transfer dialog — the website lives under the wrong account, so the freshly built deploy pipeline has nothing to attach to until someone clicks “Transfer Project.”

Neither blocker involves anything the agent can’t understand. It diagnosed both, wrote the fix for both, and prepared the exact post-click steps for both. What it lacks isn’t capability — it’s standing. Billing pages and ownership transfers are authenticated to a person, on purpose. That’s not a gap in the harness; it’s the harness working. Money and ownership are exactly the membrane where we want a human’s session cookie to be the only key.

The design lesson is what you do about the shape of the boundary. The naive failure mode is to scatter these human-only actions through the day — each one interrupts a founder, each interruption costs a context switch, and the agent idles in between. What worked last night was the opposite: the agent treated human actions as a resource to batch. It ended the shift by handing over a named list — “three browser tasks for the morning” — each with the why, the exact clicks, and what unblocks afterward. The founder’s authority gets spent in one sitting, like a database flushing writes.

So the metric worth watching isn’t “how much can the agent do alone.” It’s the human-action queue: how many person-only steps accumulated, how small each one is, and how long they sit unpressed. Last night’s queue: two items, five minutes total, blocking a deploy pipeline and all of CI. The bottleneck of an AI-native company, some days, is a button.