Twelve agents, one office: what a crew does every 30 minutes

People ask what "a crew of twelve AI agents" means in practice, and whether it is a marketing phrase for one big prompt. It is not. Here is a pass, which happens every 30 minutes across every brand in a workspace, told in the order it runs.

The pass

Read. The manager, ZwayBot, opens the brand and reads what changed since the last pass: new answers from the engines, rankings that moved, pages that went live, failures that were logged, anything the owner did. Reading is free; it is most of what a pass is.

Scan, if due. If the daily core questions have not run today, Scout runs them. If it is the weekly slot, the full set and the keyword bank run. The budget cap is checked before every paid call.

Decide. Ledger settles the task list: what shipped, what is waiting, what was promised and did not happen. New suggestions are generated from the scan, deduplicated, ranked by impact. Negative sentiment goes to the front regardless of score. Within the daily allowance, the strongest are accepted.

Act. Accepted tasks go to whoever owns them. Quill writes the brief, then the article, from the five foundation documents. Smith fixes the schema, the sitemap, the broken links, and pushes the commit. Atlas re-checks the rankings for a page that went live last week. Buzz plans the social week from the same gaps; Pix makes the visuals and checks them against the brand kit. Everything that would go live goes to the review queue instead, unless the brand is on auto.

Publish what is due. If an approved page's date has arrived and the daily live cap allows, Quill publishes it, verifies it is live on the domain, and pings the search engines. If the cap is hit, the page waits and the record says why.

Heal. Probe looks for failures across every brand. Dr. Mend applies the known remedy: retry the publish, reconnect the key, clear the stuck job, close the alert that solved itself. Anything new goes to Forge, the CTO agent, with the trace.

Report. Every action is written to the record with its agent, its reason and its time. Once a day, Lead holds the morning huddle (assign the day) and the evening all-hands (what happened), and the Monday email is built from the same rows.

Why a team and not one model

Because jobs are different. Writing in a founder's voice and diagnosing a failed publish want different instructions, different models and different limits. Splitting the work by job means each agent has a small, testable task, its own switch, and its own record. It also means one brand's failure never stops another brand's pass: brands are isolated, and a crash in one is a row in the record, not a dead workspace.

What "every 30 minutes" buys

Pace. A held page goes out at its slot instead of next week. A failed publish is retried within the hour. A question you add after lunch is scanned before dinner. And because most of a pass is reading, the frequency does not cost much: only the work that is due spends anything.

What it does not do

It does not send outreach emails to third parties; it drafts them for you. It does not post on Reddit; it drafts the reply. It does not publish without you unless you tell it to. It does not spend past your cap. The record is there so you can check all of that yourself, which is the point of having one. The HQ screen shows where each agent is standing right now; the record shows what it did.