Testing an agent before you rely on it Guide

Before you hand real work to an agent, it is worth watching it do the job once. A Test run lets you do exactly that: it runs your agent from start to finish so you can confirm it behaves the way you expect.

What a Test run actually does

A Test run is a real run, not a preview or a simulation. Your agent uses its actual tools, its connected integrations, its skills, and everything in its knowledge base, then works through the task exactly as it would on a schedule or when a trigger fires.

That is the point. Testing against the real setup is the only way to know that the agent has the right capabilities wired up and that it produces the output you need. Because it is a genuine run, the results are saved and available afterward, just like any other run.

Because the agent uses its real tools and integrations, actions such as sending an email or posting a message happen for real. If you want to avoid reaching customers or teammates while you are still refining the agent, test with safe sample inputs first.

How to run a test

1. Open the agent from your list of agents to reach its detail page.
2. Start a Test run for the agent. If the task needs input to get going, provide it when prompted.
3. Let the run complete. You can follow along as the agent works through the task.
4. Review the outcome and any files it produced.

Where the results are saved

Every Test run is recorded like a normal run, so nothing you learn from it is thrown away.

• The run itself appears on the agent's Deliveries tab, alongside runs that happen automatically.
• Any file the agent generates — a report, a PDF, an image — is saved there and also shows up under Output files on the Files tab, where you can preview or download it.

If you cannot find an output right after a test, check Deliveries first. That is where finished work lands.

What to check in the result

A good test tells you more than whether the agent finished. Look at the substance:

Did it use the right capabilities?

If the agent could not complete a step, it may be missing a tool or an integration. Confirm which tools and connected accounts it has on the Tools tab, and check that any integration it needs is actually connected.

Was the output accurate and on-brand?

Generic or off-target results often mean the agent is missing context. Adding the relevant documents, price lists, or brand guidelines to its Knowledge base grounds its work in your specifics.

Did it follow a proven method?

If you expected a structured, expert approach and did not get one, the agent may need the matching skill installed so it follows a step-by-step playbook for that job.

Refining and testing again

Testing is meant to be iterative. When something is off, you do not edit the agent's inner workings by hand — you describe the change you want.

1. Go to the Build tab and tell the AI builder, in plain English, what to change. For example, ask it to add a capability, adjust how the agent writes, or change how it notifies you.
2. Run another test to confirm the change did what you wanted.
3. Repeat until the agent handles the task cleanly on its own.

Version history keeps a record of these changes, so if an adjustment does not work out, you can roll back to an earlier version and try a different approach.

When you are ready to rely on it

Once a Test run produces the result you want, using the agent's real tools and real integrations, you can trust it to do the same when it runs on a schedule or in response to a trigger. Set up how you would like to be notified, and let it work — checking Deliveries whenever you want to see what it has done.