Workflows
Short answer: A Workflow is an ordered list of steps that runs 1 → 2 → 3, where a later step can start from an earlier step’s saved output. It’s the automation surface that carries schedules and triggers; agents don’t schedule themselves.
What is a workflow?
A workflow is a sequential list of steps you author, stored in the workspace. There’s no branching canvas and no visual graph: steps always run in the order you put them in, and any conditional logic (“if X, do Y, otherwise Z”) lives inside a step’s own instructions, decided by the model running that step.
Each step is one of two kinds:
- Have AI do something: an inline, one-off worker defined right in the step: a model, an optional tool allow-list, and a cap on tool rounds. Good for a single bounded task that doesn’t need to be reusable elsewhere.
- Run a saved agent: the step delegates to a saved Agent, running with that agent’s own model, prompt, and tools.
Steps hand data forward: give a step’s result a name, and a later step can start from it (its content gets folded into that step’s instructions automatically). A step can also write its result straight to a file, in addition to (or instead of) passing it forward in memory.
Workflows are what carries Triggers: schedule, email, Discord, file-watch, or another workflow finishing. Agents are on-demand workers; wrap one in a one-step workflow to put it on a timer or fire it from an inbound message.
How do I build and run a workflow?
- Open the Workers panel and switch to the Workflows tab, then create a new workflow.
- Add a step and pick its kind: Have AI do something or Run a saved agent.
- For a Have AI do something step, pick a model, optionally check off which tools it can use, and write its task instructions. For a Run a saved agent step, pick the saved agent and write the task it should run.
- Optionally name the step’s result so a later step can start from it, or give it an output file to write to disk.
- Reorder steps by dragging, then save. Run it with Run now, or attach a trigger so it runs on its own.
Good to know
- Steps run strictly in order; there is no graph view.
- Each step is itself a bounded worker run (the same engine an agent uses), so a Have AI do something step can call several tools and make its own decisions within its round cap before handing off.
- A step (and an agent it delegates to) can be confined to specific input/output folders; see Scoped folders.
- Every completed run gets a History card with one-click Revert that undoes every file the run touched, across all its steps; see Run history and revert.
- Off-machine sends inside a step follow the same pre-authorized-destination rule as agents: anything not on the allow-list parks the run instead of sending.