One agent answers. A crew finishes the job.
A single agent handles a single ask. A crew is a standing team that runs a whole piece of work end to end — research feeding analysis feeding a written deliverable — with each step handed to the specialist built for it, and the steps you care about stopping for your approval before anything leaves the building.
How crews works
Each step names the specialist that runs it and automatically sees what the previous steps produced. Steps you mark to run together execute in parallel; the rest run in order. That chaining is what turns several agents into one job.
A step can be required to return structured data, which is validated and retried against the schema. It can carry a guardrail in plain English — 'must include the total in dollars' — checked by a second model, with the reason fed back on the retry. A guardrail that never passes fails the run rather than shipping the output.
Mark any step for approval and the whole run pauses there. Approving continues from exactly that point — finished work is replayed, never re-run, so nothing is repeated and nothing is billed twice. Approve from the web or from your phone.
The outcomes, plainly.
- Whole jobs finished by a team, not fragments by one bot
- Structured, guardrailed output instead of whatever came back
- Nothing irreversible happens without your one tap