Workflows
How a workflow comes into existence in an interview, learns from real runs, and becomes deterministic code once you decide it is ready.
A workflow is one recurring job you have handed over. Answering inbound leads. Reconciling yesterday's invoices. Filing the weekly report. You describe it once, it runs with you watching for a while, and then it runs by itself.
Every workflow is built around an SOP, a written procedure in plain language. The SOP is the contract between you and the platform: it starts rough, gets sharper run by run, and stays readable the whole way through. The code implements the SOP. The SOP explains the code.
The arc
Setup you describe the job in an interview, then publish
Learning real runs, one at a time, with you approving and accepting
Operational codified: deterministic steps, in parallel, no model in the loopTwo stages, and one human decision between them.
Learning is agentic. A single agent works from the SOP with full flexibility, asks you before anything consequential, and gets the job done. Every run you accept becomes an example of what the procedure really looks like against real data.
Operational is codified. The same job runs as ordered steps, mostly plain code, with an agent kept only where judgement is genuinely unavoidable. A fully codified workflow costs zero model calls, which is what makes the thousandth run cheap and fast and identical to the first.
The bias runs hard toward the second one. An agent surviving into a running workflow is a place the procedure could not be pinned down, treated as a cost rather than a feature.
You decide when it flips
Nothing promotes itself. There is no counter that trips, no confidence score, no switch flipped on your behalf. The platform surfaces readiness signals, and you walk the procedure through with it and confirm at the end.
The same is true in the other direction. When a codified workflow fails, it does not repair itself and it does not quietly un-automate itself. It brings you a failure you can read and a recommendation you can overrule.
Where things live
Creating a workflow
Naming the job, the setup interview, the SOP blueprint, and what Publish actually does.
Teaching it through real runs
Sequential learning runs, accepting and rejecting, and the readiness meter.
Codification
The walk-through, what you can pin, and what changes once you confirm.
Editing and revising
Adapt or reset, repairing a failed run, and stepping back to learning.
The workflow directory
The versioned artifact: one procedure, one set of steps, one set of files, one version number.
Reference files
The price list, template or policy PDF a workflow works from.
Statuses
The workflows list shows exactly one status per workflow, and every other surface derives its badge from the same rule:
| Status | What it means |
|---|---|
| Setup | Created, never published. |
| Learning | Published, running agentically, teaching itself from your verdicts. |
| Operational | Codified, running as deterministic steps. |
| Paused | You disarmed the trigger. Publish again to bring it back. |
| Needs review | A codified run failed and something is waiting on you. |
What a workflow is not, yet
Worth knowing before you design around it:
- Steps run as an ordered sequence. There is no branching graph. Variation lives inside a step, or in an agent step that makes the call.
- One workflow, one owner. Approvals go to the person who owns the workflow. There are no approver chains, no escalation on a timeout, no shift-aware routing.
- No staging environment. A workflow is published or it is paused. Rehearsals happen through test runs against a draft, which cannot touch what is live.
- No templates. Every workflow starts from a conversation about your job, not from a library entry.
Testing a trigger
How to fire an event-driven workflow against a realistic payload before anything real reaches it, and what a test run will refuse to invent.
Creating a workflow
Name the job, talk it through with the assistant, then publish. What the setup interview writes and what Publish actually does.