Creating Orchestrations
Orchestrations are built in Settings → Orchestrations using the visual canvas.
Using the canvas
Click New Orchestration (or select an existing one from the dropdown at the top). Give it a name and an optional description in the fields at the top of the page.
Adding steps
The step type toolbar runs across the top of the canvas area. Click any step type button to add that step as a new node on the canvas:
LLM · Agent · Tool · Evaluator · Parallel · Merge · Loop · Human · Transform · Extract JSON · If/Else · Switch · Print · End
Connecting steps
Drag from the output handle of one step node to the input handle of the next. This sets the next_step_id for that step. For branching steps (Evaluator, If/Else, Switch), multiple output handles appear — one per route.
The first step you add is automatically set as the entry point. To change the entry point, right-click a step node and select Set as Entry Point.
Configuring a step
Click any step node to open the Step Config panel on the right. Fill in the fields for that step type. See Step Types for details on each type.
State schema and guardrails
At the bottom of the canvas, two tabs let you configure the orchestration's global settings:
State Schema — define the variables your orchestration uses. Click the + button to add a key, set its type and description. The schema generates input forms when running manually and documents the data contract.
Guardrails — set Max Total Turns, Timeout in minutes, and an optional Max Cost USD budget for the run.
Running from the canvas
Click Run to start the orchestration. A run input form appears based on the state schema. Fill in the initial values and click Start.
The Run Log tab at the bottom shows a live stream of events as the orchestration executes — step starts, agent turns, tool calls, step completions, and human input prompts. Each event is colour-coded and expandable.
Recent runs
The Recent Runs tab shows past run IDs with their status (completed, failed, paused). Click any run to inspect its step history. For failed runs, a Resume button lets you restart from the point of failure.
AI Builder
Click Build with AI to describe your orchestration in plain text instead of building the DAG manually. The AI Builder uses your configured LLM to generate the step structure, routing, and configuration. After generation you can edit the result on the canvas.
See AI Builder for the full walkthrough.
Other actions
Save — persists all canvas changes. Changes are not saved automatically.
Deploy as Agent — creates an orchestrator-type agent linked to this orchestration, letting you chat with it directly from the sidebar.
Duplicate — creates a copy of the orchestration with a new ID that you can rename and modify independently.
Delete — removes the orchestration permanently. This cannot be undone.