Agent workflows: AI that follows your written procedure

Until April, every workflow was TypeScript that ran the same steps each time. That suits moving data between apps, and breaks down on work that needs judgment: deciding which stories in a feed matter, triaging a reply, drafting a response. Workflows now come in two types. Code workflows work as before. Agent workflows hand each run to an AI agent that follows a written procedure, and you can switch a workflow between the types.

How it works

• Three files. The same manifest code workflows use (credentials, triggers, inputs), a plain-English runbook, and an agent configuration with the model, a time limit, and a reasoning effort.
• Scripts as utilities. The builder writes a small helper script for each external call, and the agent decides when to run them. If a script breaks mid-run, the agent can fix it and the fix is kept.
• A living runbook. After each test, the builder writes what worked (endpoints, response shapes, working code) back into the procedure, so later runs follow it instead of exploring again.
• Run history as context. Agents can pull a workflow's recent executions to debug failures against real runs.

Also shipped

• An Activity calendar of past and upcoming scheduled runs, and per-type email alerts.