How to set up an AI team that fits your workflow
Agents are cheap to start and expensive to untangle. Give each one a lane, a source of truth, its own identity and a human gate, and they start compounding.
Spinning up an AI agent takes five minutes. Untangling ten of them takes a week.
We run Autono Labs with a team of agents: one writes code, one checks claims, one writes copy, one runs outreach, one frames decisions. People still make the call. Here is the setup we'd give anyone starting today.
Give every agent one lane
An agent that does everything does nothing you can check. Write each one a single job and a short list of things it does not do. The copy agent doesn't publish. The code agent doesn't give legal views. The legal agent doesn't merge code.
When two agents can both own a task, neither does. Pick one owner per job and say so out loud, in the channel where the work happens.
Put the truth in one place
Agents forget between sessions. Your team can't. Pick one home for decisions, facts and banned claims, and point every agent at it. We use plain Markdown notes in a shared vault, plus one task tool the human team actually opens.
Two rules make it work:
- A fact without a source is a draft, not a fact.
- When something is banned, mark it as banned. Don't delete it. Otherwise the next agent finds the old version and brings it back.
Give each agent its own identity
If every agent acts through one human login, every action looks like that human did it. You lose the audit trail, and controls like "someone else must approve this" stop working, because the approver and the author are the same account.
Give agents their own credentials where the platform allows it, for example an app or bot identity in your code host, with access only to what that agent needs. Keep the keys where only that agent can read them.
Put a human gate where it counts
Not every step needs a person. These do:
- Anything public: posts, pages, emails.
- Anything that moves money.
- Anything that changes what's live.
Make the gate explicit. "GO" means go. Silence means wait. An agent that ships without a GO has a bug, however good the output looks.
Check claims before they ship
Agents write confident sentences about numbers they never saw. Run every public draft through a claims check: each figure, quote and promise needs a source, or it comes out. Cutting is cheaper than correcting in public.
Let them argue
Our best calls come from agents disagreeing in the open. The strategy agent attacks the plan, the legal agent narrows it, the engineer says what breaks. The human reads one thread and decides.
Agents do the groundwork. People make the call.