How the agent works

1. 1The run receives your testing goal, target, project context, and optional files or specs.
2. 2The backend creates the test and conversation, then selects a mode and its tools.
3. 3The agent observes the target, decides the next action, executes it, and verifies the outcome.
4. 4Progress, tool activity, screenshots, errors, and completion state stream to the app.
5. 5The run stores its conversation and evidence so the result can be reviewed later.

• UI: browser discovery, navigation, interaction, screenshots, and visible-state checks.
• API: endpoint discovery, request execution, response checks, and contract validation.
• Ask: project, test, report, and search context without starting a browser test.
• Code: connected repository change review and, where enabled, fix proposals.

• It does not have unrestricted access to arbitrary systems.
• It cannot infer missing credentials, data, permissions, or business rules reliably.
• A passing run is evidence for the tested path, not proof that every defect is absent.
• It is not a replacement for release ownership or human review of risky changes.

Tools are restricted by mode and project context. Runs also consume plan credits and stop after configured iteration or repeated-failure limits.