Overview
A dynamic workflow turns a single complex request into a structured, multi-step plan that the super agent runs end-to-end. Instead of working through a large task in one long back-and-forth, you describe the goal, the super agent proposes a set of connected tasks shown as a visual graph, and — once you approve — each step runs in turn with live progress, summaries, and results. Workflows are ideal when a request naturally breaks into parts: comparing several options, researching multiple sources in parallel, or producing a deliverable that depends on earlier steps.Dynamic workflows are ad-hoc — you describe a goal and the super agent runs it once. For a multi-step process you repeat with different inputs, an Agent packages the same kind of work into something reusable.
Starting a workflow
Type the/workflow command in the chat input, followed by your goal:
| Command | What it does |
|---|---|
/workflow <goal> | Describe a goal; the super agent proposes a workflow for approval |
/workflow run <goal> | Same as above — run is optional |
/wf <goal> | Short alias for /workflow |
/workflow | With no goal, the super agent helps you shape one |
Referencing skills with @
While describing your goal, type @ to reference a specific skill or connected service inline. The super agent factors those capabilities into the plan it proposes:
@ menu lists your enabled skills and connected services so you can pick the exact capability you want a step to use.
Approving the plan
Before anything runs, the super agent proposes the workflow as a structured block in the chat — not just text. The proposal shows:- A name and description for the workflow
- The phases and tasks it will run, with how many run at once
- A visual graph connecting the phases, so you can see how steps depend on one another
- A Needs approval badge while it waits for you
Review the proposed plan
The super agent responds with a workflow proposal — a visual graph of connected tasks plus an expandable breakdown of each phase.
Approval is required — a workflow never runs until you click Approve. If a proposal isn’t what you wanted, just describe the change or send a new request instead of approving.
Watching it run
Once approved, the workflow graph updates live as the super agent works. Each task carries a status so you always know where things stand:
| Status | What it means |
|---|---|
| Pending | Queued, not started yet |
| Running | Currently executing |
| Completed | Finished successfully |
| Failed | Hit an error |
| Cancelled | Stopped before finishing |
Controlling a running workflow
While a workflow is active, controls appear in the workflow block:- Pause — stop new tasks from starting while preserving completed work
- Resume — pick up again from where you paused
- Cancel — stop the entire workflow
- Retry — re-run an individual task that failed
Inspecting a task
Click any task in the plan to open a side panel showing exactly what that step’s sub-agent is doing — its thinking, tool calls, and generated output stream in real time, using the same rendering as the main chat. This makes a multi-step run transparent: instead of waiting for a final summary, you can open any task — running or finished — and follow its work step by step. A task that hasn’t started yet shows as queued; once it begins, its activity streams in and scrolls to the latest as it arrives.
Good to know
- You approve, the super agent structures. The super agent decides how to break the goal into phases and tasks based on what you describe — the plan isn’t hand-edited before it runs.
- Workflows run on credits, like any other session. The proposal shows an estimate, and the actual cost depends on the models and tools each task uses. Larger plans with many parallel tasks cost more.
- Shared threads are read-only. When you open a shared, read-only thread, workflow controls don’t appear — viewers can follow along but can’t start or approve a run.
- For processes you repeat, reach for an agent. A dynamic workflow runs your goal a single time. If you run the same multi-step process regularly, an Agent packages that kind of work with structured inputs you can re-run, and can be put on a schedule.