The canvas
The visual flow editor where an AI employee is assembled.
Part of the crew rollout: Agents shows as Soon in the app until it reaches your account. See Hiring your crew.
An agent is a flow, not a prompt. You build it on a canvas: drag in nodes, connect them, and the path data takes from input to output is the agent's logic. It's the same flow-graph engine the PRD journeys and Design canvas use, so it'll feel familiar.
Open the editor from the agents list (Hire, then Expert mode for a blank canvas), or edit an existing one.
Nodes and connections
Each node is a step. You drag one from the palette, drop it on the canvas, and draw a connection from one node's output port to the next node's input. Data flows along those connections in order.
| Node | What it does |
|---|---|
| Input | The entry point: the message or event that starts a run. |
| Output | The final result the agent returns. |
| Agent | The thinking core: an LLM with instructions, tools, and memory. |
| Sub-agent | Calls another agent you've built, as a step. |
| HTTP | Calls an external API. |
| App | Runs an action on a connected integration (Slack, GitHub, …). |
| Branch | Routes down different paths based on a condition. |
| Transform | Reshapes data between steps. |
| Code | Runs a custom code step. |
| Memory | Gives the agent session history to remember a conversation. |
| Prompt template | Fills a text template with values from earlier steps. |
Every new agent starts with a simple Input → Output graph; you build outward from there.
Working on the canvas
- Add: drag from the node palette onto the canvas.
- Connect: drag from an output port to a compatible input port. Incompatible connections (like wiring a memory port into a data port) are refused.
- Select, move, delete: click to select, drag to move,
Deleteto remove. - Undo / redo:
Cmd/Ctrl+ZandCmd/Ctrl+Shift+Z, with copy, paste, and duplicate (Cmd/Ctrl+D).
As you work, the editor validates the graph live and flags problems (a node with no connection, a branch with no cases) before you ever run it.
Build it by chatting
You don't have to wire everything by hand. The AI co-builder lets you describe what you want ("add a step that posts to Slack when the input mentions a bug") and it edits the graph for you, showing exactly which nodes and connections it added before you accept.