Authoring with AI

You can build and evolve workflows without opening the constructor at all — by talking to your connected AI client (e.g. Claude Code). The agent can list your workflows, read any of them, create new ones, and make targeted edits, including team and personal variants.

Migrate a process in one conversation

The most valuable use: your team already has processes — in a wiki, a checklist, a senior engineer's head. Describe one to the agent and ask for a workflow:

"Our release process: collect the merged PRs since the last tag, write release notes in the changelog, tag the version, and post a summary in the #releases channel. Create a release workflow from this."

The agent drafts the phases and steps, creates the workflow, and it's immediately visible on the AI Workflows page for you to review and refine — and immediately runnable by every agent. No format to learn, no migration project: one conversation per process.

The same works for adjustments: "add a step to our task_implement — after the tests pass, check the changelog needs an entry" is a complete request.

Safe by design

Agent edits are deliberately conservative:

• An edit changes only what it explicitly names. A step the agent doesn't mention is never touched, and deleting a step requires saying so outright — a careless or truncated request can't silently wipe a workflow.
• Every agent edit lands in the activity log with exactly what changed, attributed to you.
• Connected clients pick up the new version instantly.
• The built-in chat agents deliberately can't author workflows — this capability belongs to the local client you drive directly.

Review where you like

Authoring by conversation and editing in the constructor are the same thing at two speeds: the agent gets you a complete draft fast; the constructor gives you precise control. A common rhythm is to dictate the first version to the agent, then fine-tune wording in the constructor — checking the Agent view tab to see exactly what agents will receive.