Run a company with agents

Running a company with agents does not mean handing your business to an AI and hoping. It means the recurring work that keeps a company alive — outreach, content, support, reporting, follow-ups, the coding in between — runs as repeatable jobs that agents execute and you direct. You stay the judgment in the system. Agents become the capacity.

This section is the deep dive behind the Operate stage of Build Your Company. The pages after this one walk through where agents fit first, how you keep control without becoming the bottleneck, how autonomy and budgets work in practice, and how one installed playbook grows into a running operation. Everything here builds on the product model from How Task Machine works. Read that first if you have not.

The work that runs a company repeats

The hard part of a small company is rarely the product. It is everything around the product that comes back every week: the outreach that keeps the pipeline warm, the content that keeps you visible, the support replies, the reports, the invoices. Doing that work from scratch in a chat window every week loses the context, the ownership, and the record of what happened last time.

The shift that makes agents useful is treating that work as a system instead of a series of heroic one-offs. Each recurring job gets a definition — who does it, when it runs, what needs your sign-off — and then it keeps getting done, the same way, with your judgment applied exactly where you decided it matters.

You direct, agents execute, one inbox decides

Task Machine expresses that system through three surfaces. In Discuss you reason about what to do — this week's priorities, how to approach a goal — and that thinking fans out into concrete tasks for the agents that do them. The Inbox is where everything needing your judgment comes back: an agent's question, an approval before something irreversible, a finished draft to review. Tasks are where you steer a specific piece of work when it needs you, with the full history of what happened.

That loop — direct in chat, approve from the inbox, steer in tasks — is the whole job of running an agent company day to day. After setup, most owners spend most of their time in the inbox, acting on decisions rather than watching work happen.

Control is a dial, not a switch

Agents in Task Machine start supervised: everything consequential waits for your approval. As an agent proves itself on a kind of work, you raise its autonomy level so routine actions proceed and only the risky ones interrupt you. Budgets cap what any agent can spend in tokens or money, and verifier steps check work before it reaches you. Control is not a global on/off — it is a set of boundaries you place exactly where your judgment is required, and loosen deliberately.

The first week, concretely

That model is easier to trust once you have seen it at desk level, so here is what the first week actually looks like. Day one is setup: you install a playbook for one recurring job — weekly content, say — and it arrives as a working shape: an agent with instructions, a workflow with the approval step already placed, and documents about your voice and audience waiting for your edits. Agents run on your own machine, connected through the tama CLI, so the rest of the day is connecting that machine and adjusting the playbook's documents until they sound like your company rather than a template.

Days two and three bring the first runs. The agent works Supervised, so its drafts and questions land in your inbox instead of anywhere public. Expect to edit heavily — that is the point of these days. You are not grading the agent. You are teaching it, and every correction you make on the task goes somewhere durable: into its instructions, its memory, the documents it reads.

The rest of the week settles into the loop this page described in the abstract: the workflow runs, a draft arrives in the inbox, you approve it or send it back with a reason. Each cycle costs you less attention than the last, because the corrections stuck.

By week two or three, approvals are going through unchanged, and that record — not a feeling — is your cue to raise the agent's autonomy so routine steps proceed and only the consequential ones interrupt you. The approval step before publishing and the budget cap never moved. What changed is where your attention goes: from reading everything to deciding the few things that actually need you.

The rest of this section takes each of those ideas in turn, starting with the most common first question: which work to hand to agents first.