Skip to main content
Every project has a saved-test agent that reads your steps, drives the app, and decides whether each step passed. Out of the box it relies on general testing knowledge plus what it has learned about your app. Agent instructions let you add your own standing guidance on top: notes about how your app behaves, what to avoid, and how strictly to judge results. The instructions are injected into the agent’s system prompt on every run of the project - manual, scheduled, group, PR, and discovery runs alike - so you write them once instead of repeating them in every test.

​
Where to set them

  1. Open the project in the dashboard and go to the Settings tab.
  2. Scroll to Agent instructions and type your guidance into the text area.
  3. Click Save. The next run of any test in the project picks them up.
Clear the field and save to remove them. Instructions can be up to 4000 characters.

​
What to put in them

Good instructions describe things the agent cannot infer from a single test:
  • App quirks: “A cookie banner covers the footer on first load. Dismiss it before interacting with the page.”
  • Navigation hints: “The admin area lives under /admin and only the staff account can open it.”
  • Boundaries: “Never press Delete account or Cancel subscription, even if a step seems to ask for it.”
  • Judgement rules: “Treat any red toast or console error as a failure for the step that triggered it.”
  • Domain vocabulary: “‘Workspace’ and ‘Team’ mean the same thing in our UI.”
Keep them short and concrete. The agent reads them alongside your test steps, project memory, and saved credentials, so there is no need to restate what is already in those.

​
How they are applied

Instructions are advisory. The agent follows them on every step unless doing so would stop it from verifying what the step actually asks for, or would conflict with its built-in safety rules. They cannot make the agent skip verification or report a step as passed without evidence.

​
API and MCP

The same field is available as agentInstructions on the project API and the MCP update_project tool, so you can manage it from code or from your own AI tooling:
Send null to clear the instructions.
  • Planner instructions for the PR exploration agent steer the planner that decides what to test on a pull request. Agent instructions steer the agent while it executes steps, on every kind of run.
  • Project memory captures what the agent has learned automatically; agent instructions are what you tell it explicitly.