Creating Workflows

The Workflows list

Workflows lists the workspace's workflows, two lines each: the name and what starts it (Manual, or a schedule such as Mondays 07:00 UTC), how its last run went (when, how long, what it cost), its description, and its steps by role. Its tabs are Workflows, Schedules and Templates. New workflow opens an empty canvas; Templates and Build with AI are the other ways in.

Using the canvas

The canvas is the whole page. A bar floats at the top: the workflow's name (click it to rename), what needs fixing before it can run, Run now, help and a menu. A dock at the bottom holds Add step and Tidy up.

Adding steps

Add step, in the dock (or ⌘K), lists every step type, grouped by what each is for. Click one to add it at the next free spot, drag it to place it exactly, or type to find one by name.

Connecting steps

Drag from a step's output handle to the next step. For most steps this sets its Next step. Routing steps have one handle per outcome:

• Evaluator — one per route (add the route in its setup card first)
• If / Else — true and false
• Switch — one per case, plus default
• Loop — body adds the target to the loop's body steps; done sets where the run goes afterwards

Connecting also adds the source step's output key to the target step's input keys, if it isn't there already. Parallel branches are chosen in the Parallel step's setup card.

The first step you add is where a run starts, marked START. To change it, open a step and press Set as entry point.

Configuring a step

Click a step to open its setup card on the right — or, with the step focused, press Enter or Space; Escape closes it. Drag the card's edge to resize it. Fill in the fields for that step type — see Step Types. Problems with a step show on it, and the bar at the top counts them.

Workflow settings

Workflow settings, in the bar at the top, holds what belongs to the whole workflow:

• Description — what it does, shown on the list and to anyone who runs it.
• Schedules — the schedules that run it. See Scheduling.
• State Schema — the keys your workflow uses: each key's type, default and description. Defaults seed a run's state, and types convert what steps write. See State schema.
• Cost per run — what a run will cost, and a cap on what one may spend (below).
• Guardrails — Max Total Turns, Timeout (minutes), Stop on a state type mismatch, Result and Score runs against. See Guardrails.
• MCP clients — Offer to MCP clients, off by default: on, Claude and other apps connected over MCP see this workflow as a tool of its own.

Cost per run

The estimate says what a run will cost before it starts, from the cheapest path through the workflow to the dearest: Each run costs about 2.2–3.2 credits and $0.03 from your model wallet on Cloud, about $0.006 on your own keys self-hosted. What the workflow can't say in advance — how long a for-each loop runs, a route back, a nested workflow — reads as or more. Once the workflow has runs, the estimate uses them: Models from its last 5 runs. The same estimate is on Run now's input.

The cap stops a run that would spend more than you set. On Cloud it is Cap in credits and Cap in model dollars; self-hosted, model dollars only. Empty is no cap, the default. The engine checks before every step, in a Loop's body, a Parallel's branches and every workflow nested under this one, and a run at its cap stops before the step that would take it past: Stopped: this run reached its cap of 5 credits. It had used about 4.9 credits, and its next step costs 1. Raise the cap in the workflow's settings, then Replay. A step already working finishes, so a run can pass its cap by what its last step spent. The run's card offers Raise the cap, which opens these settings.

Under the fields, the cap is set against the estimate: Its cheapest run costs about 2 credits, so this cap stops every run before it ends, or Its dearest runs cost about 3 credits and may stop at this cap. With a cap set, the estimate adds it: Capped at 5 credits a run.

Running from the canvas

Run now, in the bar at the top, runs the last saved version; the arrow beside it takes an input for the run, which becomes user_input in state. It won't start while the workflow has errors, and while it runs, Run now becomes Stop.

Each step lights up as it runs, a person needed glows amber, and the run strip at the bottom left says where the run is. Its Run Log follows the run — step starts and completions, model and tool activity, routing decisions (one line per Evaluator decision: Evaluator chose approve (…)), warnings and questions for a person. A step that fails shows its own reason, on the step and in the log — also when the run is reopened. A question for a person is answered in the waiting step's setup card. Last run in the bar shows each step's time from the last run (and its cost, on Cloud).

Recent runs

The run strip's Recent Runs lists this workflow's last runs with their status — running, paused, completed, failed or cancelled; Runs has every run. A failed or cancelled run has Resume: for a failed run the step that failed runs again; for a cancelled one it continues from the step it stopped before. State from the steps that already completed is kept.

---

Synapse AI

Click Build with AI in the toolbar to describe a workflow in plain text instead of building it by hand. The builder plans the workflow, creates the agents it needs and saves the workflow, which you can then edit on the canvas.

See Synapse AI for the full walkthrough.

---

Other actions

Save — saves all canvas changes (Ctrl+S). Changes are not saved automatically.

Undo / Redo — Ctrl+Z and Ctrl+Shift+Z.

Deploy as Agent (Deploy as agent on Cloud) — creates an Orchestrator agent linked to this workflow, so messaging it in Chat runs the workflow. Deploying the same workflow again reuses the agent. This is how an Orchestrator agent is made: creating one through the API without a workflow is refused.

Duplicate — creates a copy with new step ids that you can rename and change independently.

Delete — removes the workflow after you confirm. This cannot be undone. A workflow something still runs is not deleted, and the refusal says what runs it: 'Invoice triage' was not deleted: the agent 'Invoice triage agent' runs it, and the widget 'Support' uses that agent. Point it at another workflow, or delete it, first.

Deploy as Agent, Duplicate and Delete are in the bar's menu (⋯). The address names the open workflow (?id=…), so a link — to a copy, say — reopens that one.