Skip to main content
The starter test is free. It reserves and spends no credits.

​
Eligibility

Each verified work domain gets one starter test. The domain comes from the account owner’s confirmed sign-up email, not the website entered for the project. Free-mail and disposable email domains do not qualify. The allowance is used when the test is queued. A failed, cancelled or deleted test does not restore it. If your starter test did not finish, contact support. Public category results do not require a verified work domain.

​
Test contents

The starter test runs:
  • One framework’s plain starter, or an empty project, based on your saved selection.
  • Each category request, or the saved job’s requests when custom-job tests are available.
  • Claude Code on Claude Haiku 5.5 and Codex on GPT-6 Luna.
  • One run per request for each agent.
A category with a hosted-service request runs six times in total. A category without one runs four times. Paid-test model choices do not change this test.

​
Starting the test

Saving onboarding setup requests the starter test when eligible. If it has not started, open Agent Picks and select Run starter test. Save a complete setup first. For an unavailable custom job, the page keeps the setup and explains why it cannot run. Each pick appears as its run finishes. Waiting, running and classifying states are still in progress. Didn’t finish means the run failed; it is not a pick against your tool.

​
Results

Read the completed count alongside the picks. The result shows how often your tool was installed, what agents picked instead, and whether they built their own solution or used a hosted service without naming a provider. Each request has only one observation per agent. Use it to inspect a choice, not to conclude that one tool reliably wins. Unfinished runs are excluded from the counts. For repeated measurements and likely ranges, request a paid test.