Steps
A Workflow's steps are what it does on every run. The Does section lists them in order, with the step kind and a summary of what each one is set up to do.
Audience: builders
Reading the Does section
Each step shows its kind and its own summary. An agent prompt shows the prompt it sends. A verified query step shows which queries it calls. You are reading the Workflow the way it will run, not a description someone wrote alongside it.
Steps run in order. If a step dies, the run fails rather than carrying on with a missing result.
What a step can be
| Kind | What it does |
|---|---|
| Agent prompt | Hands the work to an agent with a prompt, optionally with named Skills attached. This is the step that reasons. |
| Verified logic | Calls a piece of verified logic the team has approved. Deterministic, and the approved version is pinned when the run starts. |
| Verified query | Runs one or more verified queries against a Data Collection. SQL your team has approved, not SQL the agent invented. |
| Data Collection intent | Asks a Data Collection a question in your business vocabulary and lets it resolve the query. |
| Tool invoke | Calls a tool the Project has mounted, such as an integration or an MCP server. |
| Apply changes | Commits file changes produced earlier in the run into the Project. |
Choosing between an agent prompt and verified logic
This is the choice that decides how much of a Workflow is judgment and how much is guarantee.
An agent prompt handles the parts that need reading and reasoning: an email whose wording varies, a document whose layout is never quite the same, an exception that needs a call.
Verified logic and verified queries handle the parts that must come out the same every time: the lookup, the calculation, the rule. Because the team approved them and the version is pinned at the start of the run, two runs on the same input give the same answer, and the audit record can say so.
Push work down the table where you can. A Workflow that asks an agent to write SQL is harder to trust than one that calls a verified query. Where the logic is settled, make it verified logic and have the agent call it.
Skills
A step can carry Skills: named capabilities the agent is allowed to use for that step. Skills can come from the Project, the team, the organization, or the platform's built-ins.
Attaching Skills to a step rather than to the whole Workflow keeps each step's reach to what it actually needs.