Skip to main content
A discovery run lets the agent loose on your app. You supply only a goal, such as Explore the checkout flow and find bugs, and the agent decides what to test as it goes. There is no saved test, no step list, and no pull request. Use it to surface bugs your authored tests never anticipated.

​
When to use it

A discovery run does not save a test. When it finds a flow worth keeping, turn it into a saved test so it runs on every change.

​
Start from the dashboard

  1. Open your project and select Tests.
  2. Click Add test and switch to the Discovery run tab.
  3. Describe the goal. Leave it empty to use the default goal, Explore the app and find bugs. A typed goal must be at least 10 characters.
  4. On mobile projects, pick the device and the uploaded app to explore.
  5. Submit. The dialog closes and the live run opens in the Results tab.
Web projects explore the project’s URL. Mobile projects need an uploaded app before a run can start.

​
How the agent explores

  • The agent opens one exploration step at a time, gives it a title, and reports each bug it finds.
  • A failed step does not stop the run. A dead end is recorded as a finding and the agent keeps exploring.
  • The run ends when the agent decides the goal is covered, or when it hits a limit.
  • Agent instructions, project memory, and your project’s credentials apply to discovery runs like any other run.

​
Limits

Each discovery run counts as one run toward your usage.

​
Read the result

Discovery runs use a different verdict than saved tests: The run description summarizes how many steps were explored, how many passed and failed, how many issues were found, and the agent’s closing assessment. Reported bugs also appear in the project’s Issues tab. See Understand Results for steps, screenshots, and videos.

​
Start from the API

The response is 202 with status: "queued", the runId, projectId, and resolved platform. Poll GET /api/v1/runs/{id} for the outcome. See the API reference and API keys.

​
Start from MCP

Coding agents connected through the MCP server can call run_discovery with a projectId, a goal, and an optional targetUrl, then poll get_run for the verdict. Mobile projects explore the latest uploaded app.