# How Self-driving works

URL: https://testmode.ai/docs/how-self-driving-works/

> Self-driving is an agent that looks after one project's testing: it explores the app, keeps a map, plans and fixes the test suite, investigates failed runs and asks your team what only you know. What it does on its own, and where you come in.

Self-driving is Testmode's agent for one project. You give it an environment to work on and a goal in your own words; it does the testing work a QA engineer would do between releases, and hands back only what needs a person. This page explains the moving parts. To switch it on, see [Turn on Self-driving](https://testmode.ai/docs/self-driving/).

## The loop

1. **It explores.** A little every day, Self-driving opens your app in a real browser and reads it page by page. Each page joins the [map](https://testmode.ai/docs/map/): what the screen is for, a screenshot, the links to and from it.
2. **It reviews the project.** After an exploration, a deploy, a change it noticed, or when you ask, it reads the goal, the map, the test suite and the latest runs, and decides what the suite lacks.
3. **It changes the suite.** It adds the tests its explorations suggest, writes and fixes test cases, switches off the ones that no longer make sense, groups them in plans and schedules the plans.
4. **Your tests run.** The scheduled plans run like any other, alongside the runs you and your CI start.
5. **It investigates what fails.** When a scheduled, CI or plan run fails, it re-runs the failures and says what each one is: a bug, a flaky test, a test that needs updating, or an environment problem.
6. **It reports.** What needs a decision goes to the [Inbox](https://testmode.ai/docs/inbox/). What it did on its own is listed under [Activity](https://testmode.ai/docs/self-driving-activity/), each change with **Undo**.

## Explorations

An exploration is one visit to your app. It runs for about 20 minutes at most, in four parts:

- **The crawl** opens pages and reads them, following links within the environment. It fills in no form except the login, clicks nothing but a cookie banner and leaves logout links alone. Besides reading, it measures what a glance misses: error statuses, slow pages, broken images and links, script errors, text such as `undefined` on screen, figures that don't add up.
- **Journey walks** reach screens no link opens, such as the checkout after a product is in the cart, by taking a few steps from pages it read.
- **Exploratory tests** try things on the pages it read, the way a tester hunting bugs does: sort a list and check the order, search in lower case, enter an invalid value and check it is refused, reload and check a choice was kept. On a production environment that does not allow tests that change data, they stop short of anything that would keep data.
- **The analysis** compares what it found with your test suite and suggests new test cases, changes to existing ones, plans and variables, plus recommendations about defects it confirmed.

Self-driving starts its own explorations: up to the day's pages while it learns your app, then one a day to keep up. You can start one yourself with **Explore something specific**. See [Start an exploration](https://testmode.ai/docs/start-an-exploration/).

## Reviews

A review is Self-driving thinking about the project as a whole. It:

- researches the product on the web when the map does not say enough (its own site, docs, pricing, changelog) and writes a short summary of the product;
- works out the journeys users take and ranks them by your goal (critical, high, normal, low), keeping any ranking your team gave;
- marks tests that change real data, which then never run on a production environment unless that environment allows it;
- starts an exploration focused on what the map lacks, at most one a day;
- adds the exploration suggestions that serve the goal, writes, fixes and switches off test cases, and tries each new or changed test once before it joins a plan;
- groups tests in plans by purpose (Smoke for the critical journeys, Regression for the rest) and schedules them within your credits;
- asks your team what only you know, as a question in the Inbox.

It ends with an **update**: a short account of what it did and every change it made to the suite, shown under Activity.

## Investigations

When a run started by a schedule, by CI or for a test plan ends **Failed**, Self-driving looks at every failed test: what the test says, what each step did, the screenshot of the failure, the browser's logs and the test's history. It re-runs the test in a fresh browser, and files one verdict per test. For a test that needs updating, it rewrites the steps, tries them in a draft run, and offers them for you to apply when they pass. See [Investigate failed runs](https://testmode.ai/docs/investigate-failed-runs/).

## What it never does

- **It never changes a test without saying so.** Every change is a row under Activity with what was there before, and **Undo** puts it back. Choose **Suggest them** in its settings and it only suggests.
- **It never runs a data-changing test on production** unless that environment allows tests that change data. See [Pause testing and protect data](https://testmode.ai/docs/pause-testing-and-protect-data/).
- **It never fills in forms while it explores**, apart from the login.
- **It never works while testing is paused**, or while it is off.
- **It never goes past your credits.** Its levels have a known cost a month, and the schedules it plans may take at most 70% of your organization's monthly credits.

## How you steer it

- **The goal.** What matters most, in your own words. Every exploration, review and investigation reads it.
- **Answers.** When it asks a question in the Inbox, your answer is kept as a note every later run reads.
- **Notes.** Teach it about the app under **What it knows** on the map: test accounts, flows to leave alone, things that only happen in one region.
- **Journeys.** Rank a journey yourself and later reviews keep your ranking. Delete a journey or a screen that is wrong.
- **Undo.** Take back any change it made to the suite.
- **The assistant.** Ask it about the map, the Inbox or the goal, or have it start a review. See [Ask the assistant](https://testmode.ai/docs/create-tests-with-the-assistant/).

## Testers and Self-driving share the suite

There are no locks: you and Self-driving work in the same test cases. **Test cases** shows who made each test case's last change, and marks the ones Self-driving wrote this week **New**. A test case's editor lists Self-driving's changes to it, each with **Undo**. When you edit a test case, your version stands; Self-driving works from it next time.

## Related

- [Turn on Self-driving](https://testmode.ai/docs/self-driving/): Where it works, how much, and what it may do.
- [The Inbox](https://testmode.ai/docs/inbox/): What needs a decision from you.
- [Investigate failed runs](https://testmode.ai/docs/investigate-failed-runs/): Verdicts, re-runs and fixes it has tried.
- [The map](https://testmode.ai/docs/map/): Screens, journeys and what your team taught it.