Skip to main content
Before exposing an agent to real users, test it. The Test Flow drawer gives you a live sandbox chat — the same routing, the same actor turns, the same tool calls as production, but no real channel delivery happens. Outbound emails aren’t actually sent; outbound Slack messages don’t actually post.

​
Open the Test drawer

Click Test flow (the flask) in the rail on the left of the canvas. A drawer opens on the right and starts a test conversation. The Draft chip next to the drawer’s title means the test runs your latest edits, even before you deploy. Real conversations run the live version. If the test can’t start, the drawer says Couldn’t start the test and why, with a Retry button.

​
Start with Initial Parameters

If the agent has Initial Parameters, the drawer opens on an Initial parameters form. Fill in the values the test should start with and click Start test; empty fields are left out. To try other values, click Change above the chat, then Restart test. See Test a condition.

​
Send a message

Type a message and press Enter. Shift+Enter starts a new line, and the box grows as you type. The platform:
  1. Routes your message to the right actor.
  2. The actor decides what to do — reply, hand off, ask to clarify, call a tool.
  3. The reply appears in the drawer within a few seconds.
You can type your next message while the agent replies, and send it once the reply is in. Each test conversation is marked as a test in the conversation history, so you don’t confuse it with a real one.

​
Why the agent answered

Under each reply, a one-line summary says what happened, for example Pricing Specialist answered · 1 tool · 1 slot saved. Click it to see the turn step by step, in plain words:
What the lines tell you: Click a tool’s pill to open the call: the arguments it was given, what came back, and any error. At the bottom, Raw events shows the engine’s own events, for when you need every detail. This is the single most useful debugging tool in the platform. When routing goes wrong, the trace tells you exactly which decision was made and why. When a route’s condition kept the router from choosing an actor, the summary counts it (1 route hidden) and a line names it, for example Hidden this turn: Refunds — user_role is not “manager”. See Conditional routes.

​
When a message doesn’t send

A message that didn’t go through is marked under its bubble:
  • Not sent. It never reached the agent. Click Resend to send it again as it is.
  • May not have been delivered. It may have reached the agent before the error. Click Edit & resend to put it back in the message box. If the agent already replied, sending it again repeats it.
Discard removes the message and puts its text back in the box.

​
What to test

Walk through the canonical case the agent is built for. Does the right actor handle the right question? Does the conversation end cleanly?
Start with one topic, then pivot mid-conversation. Does the right actor pick up the new topic? Does the previous actor’s data persist (via slots)?
Ambiguous messages, off-topic questions, things the actors aren’t supposed to handle. Does the router escalate cleanly? Does it ask to clarify rather than guessing?
Trigger paths that should invoke an integration action. Does the actor pick the right tool with the right arguments?
In Test mode, integration actions do execute for read-only operations. Be careful with write actions on production connections — those will land in real systems. Configure dedicated sandbox connections if you can.
Provide information in one message, then ask a follow-up that depends on it. Does the actor remember? Do other actors see the same context after a handoff?

​
Iterating on prompts

Test sessions are the fastest feedback loop for prompt iteration:
  1. Open the Test drawer.
  2. Send a message that exposes the problem.
  3. Open the reply’s summary to see why it happened.
  4. Edit the actor’s Instructions. Changes save automatically.
  5. Start a new test: close the drawer and open it again, or click Change, then Restart test if the agent has Initial Parameters.
  6. Send the message again.
A test conversation runs the draft as it was when the test started. Your edits reach the next test you start, not the one already running. You don’t need to deploy to test them. To retry a real conversation, open it in Conversations and click Replay in Test. The Test drawer opens with that conversation’s first message in the box, ready to send.

​
Ending a session

Click the close button on the drawer (X). Any open session is marked ended. The session and its full transcript stay in the Conversations history, marked as a test. You can re-open it later to review what happened.

​
Test vs real sessions

The “no real channel delivery” is the key safety feature — you can fully exercise an outbound email agent without sending mail to anyone.

​
Next

Deploy an agent

Publish your edits so real users talk to them.

Conversations history

Review past conversations, including test sessions.