# Long-running work, reviews and publishing

> How slow work is tracked, and which steps always need a person.

Your AI tool can start a lot of work. Some of it takes a while, and some of it always needs a person to say yes.

## Long-running work

Builds, deployments, agent turns and playbook runs return a handle straight away: a session id, its state, a builder link and what to call next.

Your tool waits by reading the session, not by sending another prompt. Handles survive a restart or a disconnect, so nothing is lost if your tool closes.

## One turn at a time

A project runs one turn at a time. When checking what changed, believe the diff over the preview.

## Publishing

`automo_propose_deployment` only drafts a publish. It posts a "Deployment ready" card in the builder chat, and nothing ships until a person approves it there.

## Merging code

Merging code into the project branch asks for confirmation inside your AI tool. That covers approving a review, promoting on a project without review, and an admin skip. The confirmation names the work, the reviewer verdicts and any fixes not yet applied.

Confirmations expire after ten minutes. Tools that cannot ask get a link to finish in the app.

## Reviews

On projects that review builder work, a finished build parks for review instead of merging. Follow it with the review tools. See [What your AI tool can do](/mcp/what-it-can-do).

> [!warning] Auto-approve settings in your AI tool cannot bypass these confirmations.