Skip to main content

Testing before publishing

You do not have to publish to find out whether it works. Studio tests for real; it does not simulate.

Testing the whole agent​

The Test button runs the agent with the data you provide. It is a real run: durable, with a timeline, subject to the same permissions and the same limits as any other.

That is why it can serve as a decision: if you tested and it passed, you know what will happen when you activate.

Testing a single step​

A step can be tested on its own, without assembling the whole flow and without the earlier steps existing. That is what makes an incomplete draft testable — which is exactly when testing is worth something.

What the step would expect from earlier steps is replaced by pinned data.

Pinned data​

Pinned data is a sample you store in place of a step's result.

How to create itWhat happens
After a run, pinning a step's resultThat result becomes available for testing the following ones
Entering a value manuallyThe same

Rules that always hold:

  • Pinned data expires. It is test material, not configuration.
  • It is isolated per organization and per agent.
  • It does not enter the definition — publishing does not carry pinned data along.
Pinned data is stored as it came

If you pin a result containing personal data, that is what stays stored until it expires. Prefer anonymous samples.

What the test checks​

A step test goes through the full publishing validation before running. A step that would not publish is not tested with a misleading result: it is refused with the same message publishing would give.

Where to see the result​

Every test run appears in the run list, marked as a test. The timeline shows the same detail as any other: the input, each step's output, the duration, and the failure reason.

Limits​

Testing has its own rate limit, separate from the manual run limit. It exists so the test button does not become a path to generating unlimited load.

  1. Build the step.
  2. Test the step with representative pinned data.
  3. Adjust until the result is what you expect.
  4. Build the next one.
  5. When the flow is complete, test the whole agent.
  6. Activate in development.
  7. Run it for real a few times.
  8. Promote to production. See Promoting and rolling back.

Next steps​