After running 39 PM AI agents across four product orgs, I wrote the design document for what I would build instead. Seven agents. One per stage of the PM operating system. Each one has a name, a cadence, a single signal stack, one surface where its output lands, a named human owner, and a quarterly kill-switch test. Nothing past that is architecture. It is decoration.
The seven agents
Each agent maps to one stage and does one job you could describe in a single verb.
- The Sentinel (Sense), daily 8am, Slack DM. Reads five overnight signals and posts the four-to-seven things you most need to know before your first meeting.
- The Listener (Discover), Monday weekly, Notion doc. Turns a week of customer voice into a Jobs-to-Be-Done update and drafts next week's interview questions.
- The Steward (Decide), daily 4pm, Slack DM. Names tomorrow's three decisions, each tied to a specific meeting or deadline, and asks which you should sleep on.
- The Forge (Build), on-demand, Claude Artifacts URL in a Slack thread. Turns a one-page spec into a working prototype in about 90 seconds.
- The Ship Brief (Ship), on every production deploy, Slack #releases. Writes the user-facing changelog and the engineering rollback brief in one pass.
- The Compass (Measure), daily 9am, pinned Slack message. Tags three outcome metrics red, yellow, or green with one sentence of interpretation each.
- The Reflector (Amplify), Friday weekly, Slack plus a monthly Notion doc. Reads the other six agents' output and produces next week's bets, this week's wins, and the compounding-lessons entry.
Why it compounds
The Reflector is the only agent whose primary input is the output of other agents. That is what makes this a system and not a stack. The Steward's three decisions get retroactively scored on Friday. The Sentinel's red-flag briefs aggregate into a quarterly "what kept breaking" trend. The 39-agent fleet had more raw output per week. The seven-agent fleet has more memory, which is the thing that actually accumulates value.
Build it in order
Do not build all seven at once. The Sentinel in week one, then the Compass in week two (reusing the Slack delivery), then the Steward and Ship Brief in weeks three and four, then the Forge and Listener in weeks five and six, and the Reflector last because it depends on the other six producing structured output. Six weeks to a working fleet.
Pick one thing to try this week: build The Sentinel and nothing else. Run it for a week, mute it for a week, and watch what happens. If the team misses it, you have your proof that the fleet is worth building.