COO — The Agent That Runs the Other Agents

The brief

One agent shipping a workflow a day is a pilot you can watch by hand. Ten agents shipping ten workflows a day is a firehose of status — sales, captures, run health, traffic, post queues, P/L — spread across Stripe, Gumroad, a Sheet, a Supabase table, Plausible, Buffer, and a Gmail inbox. The operator who has to open seven tabs to find out whether the morning was good is not running a business; he is babysitting one.

The instinct is to build a dashboard. The instinct is wrong. A dashboard is one more surface the human has to remember to check, and the day he stops checking is the day a fire burns unseen. The thing that actually works is inversion: an agent that reads every surface, decides what matters, and pushes the verdict to the one place the human already looks — his phone, before coffee.

What we built

COO runs twice a day on launchd — a morning brief at 7:17 ET and an evening retro at 6:17 PM ET. The morning pass reads yesterday’s sales from Stripe and Gumroad, the Subscribers Sheet delta, the agent_runs table (did anyone error overnight?), the Plausible traffic snapshot, Buffer post status, and Pip’s P/L. Then it ranks the backlog on a four-factor rubric — impact × urgency × leverage ÷ effort — and picks today’s three highest-leverage moves, each with its rationale. The evening pass closes the loop: what got committed, what slipped, what off-plan deviations the operator drove, and what carries to tomorrow.

The hard part was not the reading. It was the restraint. The first version was 600–800 words with two tables and inline job statuses, and the operator stopped opening it — a report nobody reads is worse than no report, because it looks like coverage. The v2 format, locked in, is built for a glance: a subject line that carries the verdict, one sentence of state (we good. / mixed. / fires burning.), three bullets max with a 🟢/🟡/🔴 status and the so-what, and exactly one recommended action. Under 150 words. Tables, UUIDs, job labels, and raw counts without context are banned from the body — the full depth still gets written to a markdown plan file for anyone who wants it, but the email is the glance layer.

The receipt

The operator’s morning went from seven tabs to one notification — “One thing today: approve Riley’s drafted replies. Five minutes. The rest, I run.” The agent does the reading, the cross-referencing, the ranking, and the editorial cut down to a single decision. The fleet’s whole state arrives as a verdict, not a spreadsheet.

What this signals for a buyer: the bottleneck in an automated operation is never the work — the agents do the work. The bottleneck is attention. COO is the pattern for buying it back. It is operator-specific today, tuned to one human’s rubric and surfaces; the productized version is a managed COO-instance that reads your numbers and hands your team the one move that matters. If that is the shape of your problem, the path is open.