For engineering leaders who need release decisions backed by evidence, not by what the agent said. Playwright and AI agent testing, with the code and the traces published.
An AI agent says it did the job. The suite passes. The release board asks one question: how do we know? If the answer is "the agent said so", you do not have a test. You have a claim.
Each investigation takes a claim people make about AI in testing, runs a reproducible experiment, and publishes the code, the logs and the verdict. Nothing appears before it has been run.
What this means for your team: ask what the AI was given before you trust what it wrote. Tests generated from the code are snapshots of the code. Tests generated from the specification can tell you the code is wrong.
All investigationsBoth suites passed on the working app. Same model, same app, same afternoon. The suite that had seen the code let eight real bugs through, including a counter that counted the wrong thing and a delete button that deleted the wrong task.
Read the investigationTwelve checks, every finding tied to a retained trace, a verdict your release board can read without me in the room. See a sample report.
Where is delivery friction coming from, and is your AI adoption helping or hiding it? Up to eight interviews, a scored maturity view and a prioritised 90-day plan.
Migration strategy, reference implementation in your repository, CI configured for parallel runs, and a measured before-and-after on a slice of your suite.
Client work is confidential. The method is not. No logos, no anonymised stories. Every claim on this site is backed by a published experiment you can rerun, or a fact you can check on the press page.
Full details and FAQs: Playwright and AI quality consulting and training, including the one-day AI Agent Testing Workshop. Fixed quotes once scope is agreed on a 30-minute call. Remote, or on site in the UK and Europe. Quotes exclude VAT.
Teams shipping AI agents or Playwright suites who want an independent read, a published method, and are ready to show the code and the traces.
Anyone who needs the answer to be yes. The verdict is the verdict. Sponsored content is labelled and never touches a finding.
SEETEST 2026, Ljubljana. Photo: Marko Risovic.
Free videos every week on the Uncle Aaroh YouTube channel, each tool put through a real task first. Start with the guide to testing AI agents or the two-minute scorecard.
Every fact, with where to check it: press and mentions.
27 checks before you trust an agent in production: boundaries, tool-use verification, the check step, failure modes, evaluation, and the Playwright evidence to keep for each. Three pages. The same checklist used in the workshop.
Not sure where you stand? Take the 2-minute scorecard first.
Engineering Productivity & AI Quality Leader at CGI, Microsoft MVP and LinkedIn Learning instructor. Formerly Principal Test Engineer at the BBC, with 20+ years in software engineering spanning development, test automation, quality engineering and engineering leadership. Section leader on Stanford's Code in Place Python programme.
Why publish the method, the code and the traces? Because the only way to tell a tested claim from a confident one is to look at the evidence. If my own work cannot stand that test, it should not be trusted either.
Based in Manchester, UK. Views are my own. Not affiliated with Microsoft's Playwright team. Facts, mentions and appearances, with where to check each: press and mentions.
Playwright: Design Patterns
Advanced Playwright Techniques
Learning Selenium