From one playbook to an operation

Nobody sets up an agent company in a weekend, and nobody should try. The operations that work grow one job at a time: a single playbook, steered until it is trustworthy, then the next one beside it. This page traces that arc end to end, because seeing the whole shape makes each step obvious.

One playbook, one job

Start with a single recurring job — the kind of work Where agents fit first points at — and install the playbook for it: the agents, workflows, and documents for that job, bundled and ready to adapt. Resist the urge to install five at once. The constraint in the early weeks is not agent capacity. It is your attention, and one job gives that attention a single target. A content playbook, say: an agent with instructions, a drafting workflow with your approval step before anything publishes, and documents that describe your audience and voice.

Steer it for a few cycles from the inbox

Then let it run, and do your part from the inbox. The first cycles are steering cycles: drafts arrive, you correct them on the task, questions come back, you answer them, and each correction goes somewhere durable — into the agent's instructions, into the documents it reads. This is the period where the job stops being the playbook's generic version and becomes yours. You will know it is settling when your approvals start going through unchanged — the evidence trail from Deciding when to trust agent output — and that is when you raise autonomy and get most of your attention back.

Add adjacent playbooks as attention frees up

The attention you recover funds the next job. Adjacent is the right direction: if content is running, the outreach that references that content is a natural second, and then the weekly report that summarizes both. Each new playbook repeats the same arc — supervised first, steered from the inbox, autonomy raised on evidence — but each arc is shorter than the last, because the workspace the new agent joins is no longer empty.

Shared knowledge makes corrections compound

That is the quiet compounding in the system. Agents in a workspace share the knowledge library — the documents describing your company, voice, and customers — and each agent's memory carries what it learned from working with you. A correction you made to the content agent's understanding of your audience is already in the knowledge the outreach agent reads on day one. You are not training each agent from scratch. You are building one accumulating account of how your company works, and every agent added inherits it. This is why the third playbook settles in days when the first took weeks.

Agents also start contributing to the growth themselves. Through proposals, an agent that sees a recurring need can propose a task or a workflow for it — and the proposal waits for your approval like everything else. The operation begins suggesting its own next pieces, with your judgment still the gate.

A morning in the mature operation

It is worth picturing where the arc ends, because the destination is concrete. Six months in, a Tuesday morning looks like this: you open the inbox to a short queue, and every item in it is real. An approval on outreach to a segment the agent has not written to before. A question from the changelog run about whether a half-shipped feature belongs in this week's entry. An 80% budget alert on a research goal that has been digging deeper than usual. A proposal from an agent that noticed the monthly customer summary keeps being asked for ad hoc, suggesting a workflow for it. Fifteen minutes of actual decisions, and the queue is clear.

Everything that did not reach you was still checked. Verifiers held the bar on every draft, autonomy levels decided what could proceed on its own, budgets bounded what any of it could cost, and each run wrote its step-by-step history to its task — so when you want to know what happened, you look it up rather than reconstruct it. The usage view tells you what each job costs per cycle, so "what does this operation cost" is a number you read, not an estimate.

And the operation is now an asset beyond its output. The knowledge library is the company handbook you never sat down to write — voice, audience, process, corrected weekly by real work. When you eventually add a person, they inherit it the same way each new agent did. The recurring work runs on its definitions, your attention is spent where you chose to spend it, and the record of how the company operates exists outside your head.

The running workflows become the company's rhythm

At some point the arc stops feeling like adopting a tool. The set of running workflows — content on its cadence, outreach on its cycles, reporting closing each week — is the operating rhythm of the company, and your week is the triad in steady state: direction in Discuss, decisions in the Inbox, intervention in Tasks when a job needs you. The recurring work runs. You run the company.

That closes the operating-model story. To build the structure it runs on, go to the setup chapters: workspaces for the container your company lives in, and the pages that follow for members, teams, projects, and goals.