Just describe the whole thing
You don’t build the steps by hand. Describe the end-to-end task and the agent decomposes it:“Search for the top 5 AI news stories this week, write a short summary of each, and email the digest to me.”The agent plans three steps, runs the search, summarizes, then emails, passing data between steps automatically. You get a task tracker card showing each step’s status.
More example workflows
Parallel steps & data passing
Steps with no dependencies run at the same time; dependent steps wait and consume earlier results via$variables. “Get the weather in Tokyo and London and compare them”
fetches both in parallel, then composes a comparison. See
variable substitution.
Human-in-the-loop
A workflow can pause for your review before a sensitive step (e.g. before actually sending an email). You approve, edit, or cancel. Controls:Build with the wizard
Prefer a guided, conversational builder? Start the workflow wizard. It asks questions and assembles the workflow with you:
(Backed by, surfaced in the shell’s workflow builder.)
Save & schedule
Workflows are saved and can be scheduled to recur (daily digests, weekly reports, hourly checks:
Scheduling is handled by ; history lives in
your device.
Specialized roles (personas)
For richer multi-step work, the planner assigns personas to steps (researcher,
writer, coder, analyst, coordinator), so each step’s output is shaped by that role’s
expertise. A research-then-write workflow reads more like a team handoff than a single
prompt. See personas.