igergov.com
← Writing
QAProba · 7 min · BG EN ES DE

The agent writes the test. You own it.

Agents check the same thing dozens of times and burn tokens, or they write throwaway tests nobody keeps. How we got to Proba, step by step.

While an agent writes code, it often checks the same thing dozens of times: it opens the app, signs in, clicks the same buttons. Every one of those checks costs tokens and leaves nothing behind. The other option is that it dashes off throwaway tests that nobody reads or maintains, and a week later nobody remembers why they exist.

The idea behind Proba is that nothing the agent does gets thrown away. The check becomes a test. A selector it found becomes memory about the app. The sign-in is saved. A failed test becomes a ticket on the board, and the fix is recorded. Next time the agent starts from where it got to, instead of from zero.

Proba started as an internal tool at Evo Tech, and today we use it in several of our products. Every part of it appeared because, without it, something hurt.

From one recording to a test that is yours

Click “Next problem” and see how each part comes from the problem before it.

What Proba has so far
What hurts

While it writes code, the agent checks that it works. It opens the browser, signs in, clicks the same buttons and looks at the result. Then it fixes something and does it all again. For a single task this happens dozens of times, every run is paid for in tokens, and afterwards nothing is left.

What we add · The agent drives the app

The agent goes through the app once, over MCP: it opens pages, fills in fields, clicks buttons, calls the API, checks the database. Every action is executed and recorded as a test step.

proba_session_open({ appKey: "web" })
proba_start_case({ title: "Sign in with email" })
proba_act({ action: "click", role: "button", name: "Sign in" })

One test, three views

Here is the same test in all three views. Change the button text and switch between them: the change is everywhere, because all three come from the same form.

Feature: Sign in
  Scenario: User signs in with email
    Given I open "/login"
    When I fill "Email" with {{account.client.email}}
    And I press the button "Sign in"
    Then I see the heading "Dashboard"

One behaviour, three ways to read it.

Why locators decide everything

A test often fails while the app works fine. The reason is that it looked for the element by its position on the page, and the position changed.

Try it below. Add a banner at the top: the positional selector no longer finds the button, while the one by role keeps working, because the button itself is the same. Then rename the button and run the test. Now the locator by role fails too, and that is correct: the visible name really has changed. The test opens a ticket, the diagnosis suggests locators from the live page, you pick one and the next run closes the ticket.

The app
Email
••••••
Sign in
The test
1  goto('/login')
2  getByLabel('Email').fill(…)
3  getByRole('button', { name: 'Sign in' }).click()
  • form > div:nth-child(3) > spanpositional · finds the button
  • getByRole('button', { name: 'Sign in' })stable · finds the button
The board

No open tasks.

The difference is in what red means. A positional selector fails even on changes the user never notices. A locator by role fails when the user really would see something different.

When it is worth it

Proba is useful when you test your own apps and want the tests to stay with you: so you can read them, change them and run them again a month later. It handles web, API and database in one and the same test, and it has visual, layout and a11y checks.

The code is open source, under the MIT licence: github.com/evotech-bg/proba.