Welcome to the 335th issue!

More testers and engineers now build agentic QA harnesses. Naturally, this raises the question of whether AI can really replace us.

For example, Dejan Menges described his approach: The QA engineer on my laptop works nights.

Since he got lots of questions, he wrote a good follow-up: We write more code than ever. Who checks it?

His conclusion?

More on that below.

Happy testing! 🙂

Dawid Dylowicz

This Reddit discussion is a good follow-up to what I just mentioned in the Comment section above. It tells us what the testing community really thinks.

Also related:

• Why an Agentic Test Pipeline Can't Replace a Human QA by Josphine Job
• The Future of QA: AI, Automation, and Continuous Quality by Julia Pottinger

Have you seen the chart that says a bug costs 100x more in production? Apparently, the study behind that chart doesn't exist and Myke Ralduxin suggests relying on your own incident data instead.

Moreover, Martin Ivison shares a wider view on why Software isn't great, and that's getting locked in for years to come.

Will Larson describes how their AI coding agents now write a verification file for each pull request. It lists what the agent checked and what it couldn't check, which shows where to improve the setup.

Also, Betty Lin reminds us that A Skill That Runs Is Not a Skill That Helps.

Thiago Werner has a point on why AI won't speed up testing if no one checks the stories and code before.

Also, a relevant Reddit thread: How are teams keeping QA from falling behind now that devs use AI coding agents?. It continues in a sibling thread.

Banu Tutuncu did a small experiment on how many of her manual findings an AI-drafted test list would catch. There's a difference.

If you let AI agents write and merge code, make it clear which decisions are left to them and which you'll make yourself. Ayden Lee proposes the process.

On that note, someone on Reddit asked: How are you running test automation in 2026? Is anyone using AI agents?

I like to see how other companies approach testing problems. Juan Diego Campuzano Osorio shares how a bank checks its Flutter screens against approved images in CI instead of in design meetings.

Similarly, Alex Ferrer presents their story on Automating Mobile Crash Triage at ManoMano.

Deepak Sinha compares constant, ramp, staircase and spike load in k6 and explains why two tests with the same peak can end up sending very different amounts of traffic.

Additionally, Irfan Mujagić describes How to Pick Concurrent Users for a Load Test.

Is your CI/CD pipeline slow? Pradyush Badapanda advises counting how many bugs each gate catches before you cut any of them.

Moreover, María Evangelina Ferreira shares why Blocking gates got us listening to our own test results again.

A good article from John Ringler on how to build a Cypress or Playwright framework that scales. It also covers how each test can create its own data and run in any order. You can also read the second part.

Furthermore, Serhii Smetanskyi writes about The Testing Pyramid: Everybody Knows It, Nobody Follows It.

How do you make sure that AI-generated tests are correct? Omkar Mujumdar advises using a second agent to implement the bug that the new test claims to catch.

Moreover, Jérémie Fleurant asks: Who opens the report of a green run?

Dmitry Khorev describes how his screenshot runs in Maestro became the most useful test pass for the app.

Also, Ravi Rawal shares an approach to Harness Engineering for AI-Generated Maestro Tests.

Insightful lesson from Ved Prakash on testing AI agents. Before you trust a perfect score, build an agent that's meant to fail and see if your checks catch it.

In relation to that, Yesha Shah wrote a good starting point on Agentic Evals: Introduction to Autoraters (aka LLM-as-a-judge).

Here's a handy summary from Michal Drajna of the Playwright 1.65 release. The new exclusive test lock looks useful for the few tests that can't run in parallel with others.

What's more, the Playwright team shared an important Playwright and browser security reminder.

This is a helpful comparison from Shoaib Ahmed Quraishi of the latest Selenium and Playwright releases, with a fair look at what download numbers and AI agents can and can't tell us.

Similarly, Shweta Gulati describes a migration story: We Moved from Cypress to Playwright Because of Tabs.

Do you spend most of your day turning test cases into Playwright code? Vishvas Dhengula points out that AI does this well now and suggests where testers can add more value.

On top of that, Anton Gulin shares a tip on Two Playwright tests, one account. Keep their data separate if you want to run tests in parallel.

In this 15-minute recording, Karthik KK shows us how to use GitHub Agentic Workflows to run AI agents securely in CI, with an example of a Playwright test.

Standups in the age of AI... 😂

Thanks for reading!

If you enjoy this newsletter and find it helpful in becoming a better tester, please consider sharing it with others.

Dawid Dylowicz