How does a human-in-the-loop AI team actually work?
One loop, three moves: you brief the chief of staff, the ten seats produce, and everything consequential waits for your signature. Here is the whole loop, walked end to end.
Quick answer
A human-in-the-loop AI team works in three moves. You state an outcome in plain words. The chief of staff breaks it into assignments across ten specialist seats, which produce drafts, research, and reconciled work. Every consequential output queues for your approval — release it or send it back. The agents never skip the last step.
Most AI tooling asks you to become the operator: prompt, review, paste, repeat. The hybrid model flips the arrangement. The team works for you the way a staff would — briefing in, work out, decisions at the top. The difference from a normal staff is that the middle layer never gets tired and the top layer is always you.
-
Brief the chief of staff
State the outcome in plain words. Seat 01 decomposes it into assignments, sets deadlines, and routes each task to the seat that owns it.
-
Seats produce the work
Research arrives with sources. Drafts arrive in your voice. Numbers arrive reconciled. Pages arrive staged. Nothing is final yet.
-
You approve or revise
The approval queue shows each finished item with its context. Release it, or attach a note and send it back. Either way, one decision per item.
One brief, walked through the loop
An illustrative example — framed as design intent, not a shipped case. Say you run a two-person company and the goal is “get a newsletter out every Friday.”
Seat 01 (Chief of Staff) decomposes the goal: a recurring Friday deadline, a content plan due Tuesday, a draft due Wednesday, your review Thursday, send Friday 9 a.m. Seat 04 (Content) drafts Friday’s issue from the week’s notes and the topics you approved last month. Seat 07 (Research) attaches two source links behind every claim in the draft. Seat 10 (Web & Tech) stages the send in your email tool. Thursday, one item sits in your queue: the issue, the sources, the subject line. You edit a sentence, hit release, and it sends Friday at nine. Total time spent by you: the Thursday review.
What routes to the human, exactly
The approval gate is not a vibe; it is a list. Anything on it queues for you:
| Action class | Examples | Default |
|---|---|---|
| Sends | Email, quotes, proposals, support replies | Require release |
| Spends | Purchases, bookings, anything with a price | Require release |
| Publishes | Posts, pages, listings, social | Require release |
| Promises | Dates, commitments, terms on your behalf | Require release |
| Internal work | Drafts, research, summaries, maintenance | Queue for review |
The defaults are conservative on purpose. A seat can never widen its own permissions — only the human seat can, and every widening is logged.
What happens when you say no
A rejection is a note, not a dead end. Send the item back with “shorter, and drop the second section” and the seat revises against your note; the next version arrives with the note attached so you can see it was honored. Over weeks those notes become the seat’s instructions. That is how a generic team becomes your team without you ever writing a prompt.
Where the model draws its line
Honest boundaries: Hybrid V10 is a concept in development. The loop above is the design the build is implementing and documenting on the journal — not a product you can subscribe to today. For the design reasoning behind ten seats rather than a toolbox, see the team page; for the trust argument, human in the loop, explained is the short version.
How it works, in questions
How much time does the human seat take?
By design, the review queue is the only required touchpoint. Everything else — chasing, drafting, reconciling — happens before the queue. The model is sized so a disciplined review session is a real work pattern, not a full-time job.
Can an agent act without approval in an emergency?
The design answer is no. Consequential action is the human seat’s jurisdiction; a model that grants itself exceptions in emergencies is a model with a marketing department, not an approval gate.
What if I want stricter or looser defaults?
Permission classes are yours to tune — you can widen “internal work” to require release too, or narrow specific repeat sends to auto-release. Changes are explicit, logged, and reversible.
Meet the seats that run the loop. The full roster — all ten seats plus the human one — is on the team page. Building something similar? Join the interest list.
The ten seats