Testing an agent

Part of the crew rollout: Agents shows as Soon in the app until it reaches your account. See Hiring your crew.

Before an employee runs on real events, you try it in the test panel. It runs your current, unsaved graph, so you can iterate without publishing a thing.

The test panel

Open it from the editor toolbar and send the agent a message, the same way its trigger eventually will. The agent runs and the response streams back token by token.

Watching it run

The canvas comes alive while a test runs, so you can see the agent's reasoning rather than just its answer:

• The active node lights up as execution reaches it.
• Connections animate as data flows between steps.
• Tool calls and results show inline: what it called, with what, and what came back.
• The agent's thinking is visible where the model reasons before acting.

If a node errors, it's flagged on the canvas with the message, so you know exactly which step to fix.

Scenarios

For agents you'll tweak repeatedly, save a scenario: a named test input you can replay after each change, so you can confirm a fix without retyping the setup every time.

Multi-turn

A test conversation keeps its session, so follow-up messages carry context, the same way a real conversation with the deployed employee will.