Customizing Workflows

Open AI Workflows in the sidebar. The left list shows every workflow in the workspace; selecting one opens the constructor.

The constructor

A workflow is edited in place, phase by phase:

Phase cards — Preprocessing, Processing, Validation, Postprocessing — each holding its ordered steps. Add a step, edit its title and Markdown instructions, drag to reorder, or toggle a step off to remove it from the process without losing it.
Validation sits between doing the work and finalizing it. Put your checks there ("run the tests", "verify every field is set") — if a check fails, the agent loops back, fixes, and re-validates. Leave the phase empty for workflows that need no checks.
Arguments describe what the workflow needs as input (e.g. a task key) so agents know what to ask for.
The Agent view tab shows the exact document a connected agent receives — the rendered result of your phases and steps. No guessing about what the AI is actually working from.

Saving publishes immediately: connected clients pick up the new version live, and the change lands in the activity log.

One workflow, three levels

The same workflow can exist at three scopes, and an agent always gets the most specific one:

Scope	Who it applies to	Typical use
Default	Everyone in the workspace	The baseline process
Team	Members of one team	Backend implements differently than Frontend
Team · member	One person in that team	A personal variation on the team's process

Create an override from the workflow's menu, pick the team (and optionally the member), and edit the copy — the default stays untouched. Deleting an override falls back to the next level up.

Built-in defaults and your edits

The three built-in workflows ship as system defaults. While untouched, they're kept current automatically as the product's canonical processes improve. The moment you edit one, it becomes your workspace's version: yours to own, never overwritten. There's no merge to manage — a customized workflow simply stops tracking the shipped default.