Skip to main content

Agents

Agents help repeated work continue when the owner steps away. A useful agent is not a vague instruction. It is a visible workflow with an event, context, decision, tool, limits, and any approval the action needs.

A bounded workflow moving from event through context, decision, approval, and tool.

Build an agent​

  1. Open Agents in the relevant business context.
  2. Choose Create agent to open an empty canvas.
  3. Drag an event onto the canvas.
  4. Add the context or condition the decision needs.
  5. Add a decision and define its limits.
  6. Add a tool or notification action.
  7. Connect compatible ports in the order the work should flow.
  8. Add approval before a sensitive action.
  9. Save and test before launch.

Work with the canvas​

  • Drag a component from the component rail onto an empty area.
  • Drag from one connection point to a compatible connection point.
  • A valid target accepts the connection; an incompatible target is rejected.
  • Select a node to expand only the inputs that node needs.
  • Delete a selected node from its node action.
  • Use zoom controls for the whole canvas.
  • Scroll inside an expanded node when its fields are longer than the node.

Test safely​

The test environment uses the same workflow shape as the agent. Select an event with available sample context, run the test, and watch the active path move through the connected nodes.

If test data is unavailable, choose an event with existing business activity or create a safe sample record. A failed test must not silently perform a live action.

Authority and approval​

Use approval when an action could affect:

  • payment or payout details
  • a customer-facing message with material consequences
  • credit, invoice, refund, or pricing decisions
  • credentials, members, or access
  • irreversible business records

Approvals appear through notifications so the reviewer can open the exact context. Authentication confirms identity; it does not replace business approval.

Before launch​

  • Confirm the workflow has one clear start.
  • Confirm every branch ends intentionally.
  • Test success, rejection, missing-data, and timeout paths.
  • Set the narrowest useful authority.
  • Confirm the owner knows how to pause or stop the agent.
  • Review results after real events begin.