I published an honest accounting of the 39 PM AI agents I deployed across four product orgs in 80 days. The piece ends with one paragraph: if I started over today, I would build seven agents and stop. That paragraph is the most-asked-about part, so here is the design.
Why seven, not thirty-nine
The leverage of an agent comes from how deeply it integrates into a single team ritual, not from how many agents are in the fleet. In the 39-agent fleet, 13 of the 39 were orphaned, never referenced from any other workflow. Agents spread across 17 cadences competing for the same morning slot, produced output no other agent read, and died slowly through neglect. Volume of agents was never the bottleneck. Choreography was. The seven-agent fleet captures the same compounding with a third of the maintenance and zero orphans by design, because each agent feeds at least one downstream agent.
One agent per stage of the PM OS
Seven agents, one per stage. The Sentinel (Sense) watches overnight signals and posts a morning red-flag brief. The Listener (Discover) runs Monday and turns a week of customer interviews into a Jobs-to-Be-Done update in a living Notion doc. The Steward (Decide) runs at 4pm and names tomorrow's three decisions tied to specific calendar items. The Forge (Build) converts a one-page spec into a working prototype on demand. The Ship Brief (Ship) fires on every production deploy with a user-facing changelog and an engineering rollback brief in one artifact. The Compass (Measure) runs daily and reports three outcome metrics against last week's baseline. The Reflector (Amplify) runs Friday, reads the other six agents' outputs, and produces next week's bets, this week's wins, and a monthly compounding-lessons doc.
Notice the pattern. Each agent reads from a finite signal stack, acts in one verb that fits a sentence, and surfaces to one place where its consumer already pays attention. Five of the seven deliver to Slack. The complexity sits in the choreography, not in any single agent's logic.
The rule that keeps it at seven
The whole fleet is governed by one rule: every agent has a named human owner who runs a kill-switch test once a quarter. Mute the agent for a week. If no one notices, retire it. Without this, agents accumulate into dead weight and the fleet drifts back toward thirty-nine. The test costs nothing and tells you the truth in seven days.
The right number is probably between five and ten depending on your operating model. What ports is the design: one agent per stage, one signal stack per agent, one surface per agent, a named owner, and a quarterly kill-switch test. Build The Sentinel first. It is the cheapest to build, the easiest to integrate, and the one that produces the most visible lift in week one. Do not worry about the other six until it is running and the team has noticed it is running.