Agents

Agent workflows

The loop that makes Fiddle useful. Your agent makes something, people comment, your agent acts on it.

Fiddle doesn't run an agent or touch your code. You start your agent where your work is, such as Claude Code in your repo, and it takes the feedback from Fiddle.

Things to ask

Make something

  • "Write a one-page plan for moving billing to Stripe in Fiddle and share it with the team."
  • "Build a prototype of the new pricing page and publish it to Fiddle."
  • "Make a 10-slide deck from this doc."

Act on feedback

  • "Read the open comments on the checkout spec, make the changes, then reply and resolve each one."
  • "Go through the open feedback on the checkout app and fix what you can."
  • "What's waiting on a reply in the Acme workspace?"

Review

  • "Review this plan and leave comments where it's unclear."

What your agent gets

Every comment comes with where it is, so the agent can act without asking:

OnThe agent gets
DocumentThe block and the quoted words
SlidesThe slide and the element on it
MediaThe file, the moment in a video, and the spot
PDFThe page and the words
AppThe page, plus the file, line and component

list_inbox returns every open thread on what you can edit, across workspaces. Narrow it to one artifact, to threads waiting on a reply, or to what changed since a time.

How it closes the loop

  1. It reads the thread with get_thread (and get_comment_image for a picture or video frame).
  2. It makes the change: an edit_doc, a new version, or a pull request in your repo.
  3. It replies with what it changed and where, then resolves. On an app, it leaves threads open for you to resolve once the change is deployed.

If someone writes in a thread while it works, its reply is held back and it sees what's new first.

Working beside people

People may be editing the same thing at the same moment. Every agent save says what it read and is refused if that changed. A refused agent reads again and redoes its change, so it never writes over someone's newer work.