Your agent goes through your app once, checking each step along the way. Argent turns that journey into a simple YAML test that automatically runs on every pull request.
Every time an agent checks a feature, it figures out the way through your app: which screen to open, what to tap, and what should show up next. Usually, all of that is gone when the session ends.
Argent E2E keeps it as a test, so next week's branch gets the same check without anyone writing it again.
One command installs Argent, its MCP server and the skills your agent needs to record E2E tests.
npx @swmansion/argent init
Your agent records the test on a simulator, you review it as a short YAML file, and CI runs it on every pull request.
You describe the flow, and your agent performs it on a simulator or emulator one step at a time. A step is saved only after the agent sees the expected result on screen.
At the end, the whole test is replayed once to confirm it passes. Elements are found by their ID or label, not their position on screen, so layout changes won't break the test. If an element has neither, your agent can add an ID to your code for you. Tapping by screen coordinates is only the last resort, and Argent warns you when it happens.
Each test is saved as a YAML file in .argent/e2e/, so it's easy to review and to fix by editing a single line. It isn't tied to a specific device and runs on any simulator or emulator.
Besides taps, a test can wait for a screen, check text, compare screenshots, dismiss pop-ups like cookie banners when they appear, reuse shared steps such as logging in, and run a script to seed test data.
argent e2e run replays the saved steps directly, with no agent or AI model involved. If a step fails, your CI job fails too, and you get a screenshot of the device at that moment plus a command that re-runs only the failing test.
We moved the Maestro suite of a recipe app to Argent E2E. An agent did the whole migration overnight and opened one pull request by morning. On the same simulator, build and network, the 53 tests both suites share ran 1.95x faster.
Argent E2E talks to the app on the device, so it doesn't matter which framework drew the screen.
One command sets everything up: Argent itself, its MCP server and the skills your agent needs to write E2E tests. It works with Claude Code, Cursor, Codex and other coding agents.
npx @swmansion/argent@latest init
Tell your agent what to check in plain words, or paste a ticket and let it turn the acceptance criteria into a regression test. You need an agent to record a test, but not to run it.
Record the path from the start screen to the checkout summary as an E2E test named checkout.
In our GitHub Actions example, building the app, booting a simulator and running the suite takes about five minutes on a hosted macOS runner, and most of that is the simulator starting up. Android runs on a Linux runner with KVM. If you'd rather skip macOS runners, --platform ios-remote runs your iOS tests on Argent Cloud.
argent e2e run .argent/e2e --recursive --platform ios
Argent E2E checks your app the way people use it, so a broken login or checkout fails in the pull request, not in production. With Argent, your agent records them for you, so they finally get written.
npx @swmansion/argent init
Deep dives and guides from the team building Argent.