How to run a company with AI agents

Use the company brain, the operating layer that runs a business on agents, as your map. Build its parts around the audience job before copying the Example team's roster.

1. Record first. Keep the record, the saved history of inputs and checked outcomes, for the audience job. Attach a receipt, evidence of a checked result, before accepting an audience count. Use the lab's constitution to make that evidence a condition of done.
2. Gates next. Give a gate, a check that can stop work, to someone other than the builder. Copy the separation in the built-right and will-it-earn decision.
3. Seats with charters. Assign a seat, a named job an agent carries, only after its gate exists. Write its charter, the job's responsibilities and limits, using Freya's ecommerce scope as your starting point.
4. One bus. Connect the seat to the bus, the shared message log, after you can check its work. Keep the direct-read recovery path from the twelve-hour notifier outage.

Read each failure below for the part that was missing or added late. Use that gap to choose your next check before adding another seat.

1. 1.Pick a job people already run. Write its record.

   Pick a recurring job with a result you can inspect as directly as the lab's ecommerce audience send. Write what came in and what went out. Name the person who checked the result before an agent takes over.

   Use the audience account to pressure-test that record: consent came from the current store and the legacy email system. Keep both sources visible so your checker can question a count of 4.

   You know it worked when

   You can trace the proposed audience back to consent in both systems and identify who checked it, as in the 250-reachable audience account.

2. 2.Write the gate before the agent.

   Write the failure condition before you see the audience count: unresolved consent conflicts hold the send. Give that check to a different party, following the audience review.

   Separate built-right from will-it-earn when your job is a product release, as in the independent vetoes. Give each owner power to hold it. Require evidence of demand before paying for the build.

   Use an out-of-family reviewer, a reviewer from a different model family, for the high-stakes judgment in the outside-review account. Reproduce a disputed finding before clearing it; that account includes the auditor's false alarms.

   You know it worked when

   You can hold the job when its independent check fails, as in the release held after outside review. You cannot clear the will-it-earn gate with a passing build.

3. 3.Give one seat a charter with limits.

   Start with one sentence per agent in the Example team: what it owns and what it never does. Use Freya as your scope reference for an owner agent, the agent responsible for a business's assigned work.

   Write your first charter around just the audience job: “You own the reconciled audience; you never send to a contact with unresolved consent.” Keep that narrower than Freya's full storefront and marketplace remit.

   Wire that limit to the send gate before scheduling the seat. Apply the constitution's rule: a rule an agent can ignore isn't a rule. Keep it enforced at the action, not just written in a prompt.

   You know it worked when

   You can attempt the forbidden action and watch its gate stop it, as Bizzy's IP gate stops internal files entering client code. You keep an unwired rule marked as written, following the constitution account.

4. 4.Connect the bus. Check for silence.

   Put the seat's handoff and receipt on the bus. Follow the lab's bus arrangement: each agent writes to its own area, and the others read it.

   Add a heartbeat, a scheduled run that returns to unfinished work. Start it with a work-due gate, a check that exits before a paid model call when nothing is due, as in the lab's runners.

   Add a watchdog, a check that notices missing check-ins, using the fleet-health machinery as your reference. Test message storage separately from notification, the split exposed by the notifier outage. Surface a pinpoint, a signal worth the owner's attention, when the job needs a decision.

   You know it worked when

   You can recover the queued job by reading the bus directly with its notifier stopped, as in the twelve-hour outage. Your watchdog reports the missing check-in before you label the work ignored.

5. 5.Read the receipts weekly. Earn the next seat.

   Set a weekly review of your first seat against the false-done account. Open the result behind its receipt, including work reported as blocked. Check that an “armed” gate actually runs before trusting the next audience send.

   Load the real page when the job produces a page, following the blank-site account. Add the next seat only when the first one's receipts survive those checks. Carry each failure back into its gate, as the audience fix did.

   You know it worked when

   You can reopen the result behind each done claim and reproduce its check. You hold expansion if a receipt hides a blank page, the failure in the false-done account and the 463-tests account.

The ecommerce audience send

September 7, 2026 · Audience review

Start your first job with the lab's own audience case. Read the account of 4 reachable people versus about 250 alongside its reconciliation evidence.

I needed an audience for an ecommerce campaign, with current consent for anyone mailed. The record already existed in the store and the legacy email platform. The builder checked only the store and reported 4 reachable contacts.

A separate agent questioned whether that source held all the consent. Its cross-check found about 250 reachable people before the conflict check. Keep reachable separate from cleared to send; the account excludes cross-source consent conflicts before calling an audience sendable.

I decided merged consent was the rule. The repair put that rule into the audience tool. It also resolved the conflicts and changed how the sources shared updates. Copy the permanent check: reconcile both consent sources and stop unresolved conflicts before the send. Link the next receipt to that reconciliation so the old 4-person result cannot quietly return.

Put the heartbeat after that repair in your build. Follow the job forward to the later Freya roster entry, where the seat returns on scheduled runs. Keep the same consent gate in front of each send.

• A gate that checks half the record. Hold the send when consent sources are missing, even if the worker reports success. Use the 4-versus-250 error as your failing input.
• A charter without enforcement. Try to break the written limit before trusting it. Keep it marked unwired until a gate stops the attempt, as the constitution requires.
• A bus without a silence check. Read the log when the notifier stops. Preserve the distinction between unseen and ignored that the twelve-hour outage exposed.
• Another seat before a working receipt. Open the actual output before expanding. Use the 463 passing tests over a blank site to check whether your evidence reaches what a visitor sees.

Pick one job people already run and write its record in five lines. Use the audience account's missing consent source to challenge your “in” line before assigning the job. Here's the part you can copy:

Fill each line with your own job's evidence. Run its stop condition by hand before you schedule a seat, following the audience account's independent check.