The canvas
The canvas is where you build an agent. It has three regions: the node palette on the left, the graph in the middle, and the config panel on the right.
Building a graph
- Add nodes by dragging from the palette onto the canvas.
- Connect nodes by dragging from one node's output handle to another's input handle. The connection defines execution order.
- Configure a node by clicking it — its settings open in the right panel, which adapts to the node type.
- Save with ⌘S or the Save changes button.
Execution follows the graph's dependency order, not the order you added nodes.
Starting from a pattern
Rather than an empty canvas, you can start from a built-in pattern that arrives pre-wired:
| Pattern | Shape |
|---|---|
| ReAct | Agent reasons, calls a tool, observes the result, repeats |
| Chain of thought | Sequential reasoning steps before a final answer |
| Self-reflection | Generator produces output, critic evaluates, loop until a quality gate passes |
Others cover supervisor/worker splits and specialist routing. Pick one when creating the agent; you can restructure it freely afterwards.
Review before testing
The Review button in the toolbar checks the graph for problems that would break a run — unconnected nodes, missing model selection, an agent node with no prompt, a tool node with no tool chosen.
Each finding links directly to the node that needs attention, so you fix issues before spending a run rather than debugging a failure afterwards.
Testing
Test opens the runner. Provide an input (JSON or plain text, matching what your Input node expects) and watch execution stream through the graph:
- Per-node status as it starts and completes
- Input and output of each step
- Token counts and duration
- The exact node and error message if something fails
Runs are queued and executed by a worker, so the panel reflects real production execution — not a simulated preview.
Publishing
Open Settings from the agent toolbar to publish. A published agent gets a public slug and can be called through the API with a scoped key. Unpublishing revokes API access without deleting the agent.
See Run an agent via API.