TesterArmy Sends a Small Army of Agents to Break Your App First

In the spring of 2026, a 19-year-old named Szymon Rybczak boarded a flight from Krakow to San Francisco. He was not going on holiday. He and two colleagues from a Polish engineering shop called Callstack had convinced Y Combinator that they could fix the part of software that everyone hates and nobody wants to own: testing. Their company, TesterArmy, does one deceptively simple thing. It sends artificial-intelligence agents to use your app the way a stranger would, and it tells you what breaks.

The premise starts with an observation most engineers will recognize. Over the past few years, AI made writing code dramatically faster. Autocomplete became autopilot. What did not get faster was checking whether any of that code actually works. A developer can now generate a feature in an afternoon and still spend two days clicking through it by hand, or wiring up test scripts that snap the moment someone renames a button.

TesterArmy's founders looked at that gap and decided it was the whole company. "Testing is still far too painful," they wrote when they launched. Their fix is to stop asking humans to write scripts at all.

01 / THE PRODUCTTests written in English, run by a robot

Here is how it works in practice. You point TesterArmy at a URL or hand it a mobile app binary. Then you describe what you want checked in ordinary language: "Sign up with a new email, add two items to the cart, and confirm the total is correct." There are no selectors to hunt for, no XPaths, no waits to tune.

Behind that sentence, an agent launches a real browser or device. It reads the screen using vision models combined with the app's accessibility data, decides where to click, types like a person, and evaluates whether the result matches what you asked for. When it finishes, you get screenshots, a video recording, and a bug report you can act on. The messy middle - login pages, OAuth pop-ups, one-time passcodes, even CAPTCHAs - is handled for you.

One detail says a lot about how the team thinks. Every test agent gets its own email inbox. It sounds trivial until you remember that a real user signs up, waits for a verification email, clicks the link, and only then gets into the product. An agent without an inbox can only test half the story. TesterArmy gave the robots mailboxes so they could test the other half.

02 / THE WORKFLOWIt lives where developers already work

A testing tool only helps if people run it, so TesterArmy planted itself inside the daily loop. It installs as a GitHub App and runs automatically on every pull request. Before code merges, the agent reads the changes, builds a test plan from them, and posts pass or fail results - with recordings and failure context - straight into the PR as a comment. Nothing to remember, nothing to trigger.

For everything else there are more doors: a command-line tool (published openly on npm as testerarmy) for local and CI runs, an API and webhooks for custom pipelines, and scheduled runs that watch production around the clock. When something breaks, the alert lands in Slack or Discord. Integrations with Vercel and Expo let tests fire on new deploys and app builds.

These are not hypothetical demos. According to the company, its early customers found these bugs in real products, sometimes within the first couple of weeks of turning it on. The bugs were already there. The difference was that something was finally looking.

03 / THE DIFFERENCEWhy not just use Playwright?

The obvious question from any engineer is why this beats the tools they already know - Playwright, Cypress, Selenium. Those frameworks are powerful, and they are also famously brittle. A test written against a specific button breaks when the button moves. Teams end up maintaining a second codebase of tests that ages worse than the app it guards. The founders put it bluntly: tests do not scale as well as code.

TesterArmy's argument is that an agent reading the screen with vision does not care that a button moved four pixels or changed its label. It adapts, the way a person would. That flexibility is also the hard part - AI agents can be slow, unpredictable, and expensive. So the interesting engineering is underneath.

Directional only - based on the company's stated design goals, not a benchmark.

To keep runs stable and affordable enough to fire on every pull request, TesterArmy caches the "trajectory" of past successful runs so the agent does not rediscover a flow from scratch each time. It pairs vision with accessibility APIs to move faster than a vision-only approach, uses Gemini Flash as its primary model with a GPT fallback, and caps how many visual analysis calls it makes per step. The result is testing cheap enough to be routine rather than a special event.

04 / THE TEAMFrom React Native core to Y Combinator

The credibility here is specific. Rybczak, the CEO, worked on React Native infrastructure and on-device LLMs at Callstack, which is directly relevant to testing mobile apps. His co-founder and CTO, Oskar Kwaśniewski, is a React Native core contributor with more than 100 merged pull requests and a handful of open-source libraries that have been downloaded over 2.5 million times. The third co-founder, Piotr Matyjasik, rounds out a team that met building mobile tooling before they built a company around testing it.

That background matters because most AI-testing startups only do the web. TesterArmy tests iOS and Android too, which is harder and where the founders happen to be experts. It is a small team - four people - moving fast on a problem they lived through.

"Master class in onboarding... first e2e test ran in under 2 minutes."

Dima Grossman, Novu

"Running it alongside manual QA has been a huge timesaver."

Juan Di Toro, HireVoice

05 / THE BUSINESSTrying to be the Vercel of testing

Ask the founders what they want to become and they reach for a comparison every developer understands: the Vercel for testing. What they mean is a product you adopt in minutes without a sales call, that works out of the box, and that quietly becomes infrastructure you stop thinking about. No six-week onboarding, no procurement gauntlet.

The pricing follows that logic. There is a free tier to try it, a "Startup" plan that the company calls its most popular, and enterprise pricing that lands around $300 for 1,000 test runs. You start without a credit card. The model is usage-based SaaS - the more your agents test, the more you pay - which lines up neatly with the promise that testing should be something you do constantly, not occasionally.

The market they are entering is crowded and getting more so. Script-based incumbents like Playwright and Cypress own the habits of most teams, and a wave of AI-testing rivals - QA Wolf, Momentic, Reflect, mabl and others - are chasing the same idea. TesterArmy's wedges are the natural-language interface, genuine mobile support, and an onboarding fast enough that a skeptical engineer can see value before lunch.

Whether TesterArmy becomes the default or one of many depends on questions no launch can answer yet: how reliable the agents stay as apps grow gnarly, how the costs behave at scale, and whether "testing in English" holds up for the truly weird edge cases that QA engineers exist to find. For now, the bet is clear and the early signal is real - 30-plus teams running it every day, and a founding team that knows the plumbing well enough to make AI testing boring, which in this context is the highest compliment.

If the last decade of developer tools taught anything, it is that the unglamorous, repeated task is the one worth automating. TesterArmy picked the task everyone skips and is trying to make it disappear into the background. Ask again in ten years whether it worked. The interesting part is that the question is now worth asking.