Creating and Running a Task
Create one
Open Tasks and click Create Task.
Give it a title, then write the description — this is the brief the agent works from, so it's worth spending a minute on. Pick the repositories it should work in and the branch each should start from, and choose the agent and model.
Directly under the description, Acceptance criteria (optional) takes one criterion per line: "One per line. Write things you could check by looking at the site — QA checks each one." After review, QA checks each one on a preview of the change — see QA on a preview. QA checks only what's in this box, never criteria written into the description. You can change the criteria later with Edit. A Task converted from an Assistant session gets its criteria from the session's spec.
Click Create. It starts as soon as a slot is free.
Tick After the PR is merged, deploy to production to have Deployment.io deploy the change to production once you merge it. See Deploy After Merge.
Writing a good description
This is the single biggest lever on whether you get back something you can merge. The agent can read your code, but it can't read your mind.
Say what should change, where, and what "done" looks like. Mention constraints it should respect and anything you explicitly don't want touched. A specific brief is usually the difference between one round of review and three.
If you'd rather work it out interactively, start an Assistant session and convert it — you'll get a far better brief than you'd write cold.
Writing acceptance criteria
Write things you could check by looking at the site: "the pricing page lists three plans" or "the sign-up button is visible on a phone". A criterion QA can't check by looking — one that needs a sign-in, say — comes back as not checked.
Start one from your backlog
Tasks converted from a session wait in Backlog until you're ready. Press Start, confirm the model, and it joins the queue.
Watch it work
The board groups Tasks by where they've got to; switch to List if you prefer a table. For what the columns and tags mean, see The Board. Click any Task to open it.
Inside you'll see how far along it is, each repository with a link to its pull request once opened, and the full agent log if you want to see its reasoning. Repositories showing No changes are ones the agent decided didn't need touching.
Stop, re-run, delete
Stop halts a Task that's running or waiting. Anything already pushed by an earlier attempt stays — nothing is undone.
Re-run gives it another go, and lets you tell the agent what to do differently first. That feedback is the useful part: "you missed the error case in handler.go" gets you much further than re-running blind. New work goes onto the same branch, so an open pull request just picks it up.
Once any of the Task's pull requests has merged, Re-run is disabled: "This Task's change has merged (<repositories>), so it can't be re-run. Start a new Task for follow-up changes." Before a merge, and after a pull request was closed without merging, you can still re-run it.
Delete removes the Task and stops any work in progress. Branches and pull requests already pushed to your repository are untouched.