Why it mattersThe usual trade with agent-run browser tests is that they are expensive on CI and brittle on redesigns. A cache that is invalidated on UI breaks goes at both.
A browser test driven by a language model is slow and expensive on CI. The same test written as explicit Playwright steps is cheap and fast, until the next redesign. Bug0 has released a tool that tries to hold both at once. The first run uses a model. Later runs replay the steps from a cache. The model is called again only when the UI changes.
Bug0 says the project is called QA Agent. InfoQ wrote about it on 29 October 2026, in a piece by Bruno Couriol. The Bug0 landing page at ai.bug0.com credits the project to the team behind Hashnode, Fazle Rahman and Sandeep Panda. It names Accel, Guillermo Rauch of Vercel, and Naval Ravikant as backers.
How the cache works
InfoQ reports the shape plainly. A developer writes test steps in English inside a Playwright test file. On the first run, the agent reads the current page and figures out which real browser actions match the step. When a step succeeds, the system compiles that step into concrete browser operations. Those operations go into an in-memory Redis cache.
Later runs read the cache first and act on the stored operations. If a cached action fails on a changed page, the system invalidates the cache for that step. It then calls the agent again to find the new path. The new result is written back.
What else it does
InfoQ lists three more pieces. Assertions run through more than one model. Claude and Gemini each judge the result independently, and a third model is called in to break a tie. The agent can extract one-time passwords from Mailsink or Mailinator for login steps. For transient defects, a vision model watches the recorded video.
What it costs and what the licence says
Bug0 says QA Agent is free as a library installed through npm. The managed version, Bug0 Managed, starts at 2,500 dollars a month and includes an engineer who writes and triages tests. The project is not open source. The company has released it under the Functional Source License, which carries non-compete restrictions during the first two years.
Why the cache is the point
Teams that have tried to run model-driven browser tests on every pull request already know the cost. A dozen such tests can take long enough to crowd out the rest of the pipeline. Freezing the steps after the first run is the shape that makes the pattern affordable. Automatic cache invalidation on UI change is the shape that keeps it from going brittle.
The licence is the part a buyer will want to read closely. Source-available with a two-year non-compete is a long way from Apache or MIT. A team using the library should use it as released, rather than fork it. The judgement for a reader is simple. Does your UI change often enough that the cache keeps paying for itself? And is the model behind the cache lookup one you trust in six months?
Source
- Bug0 Releases QA Agent to Automate Regression Testing With Cached Playwright Replays, InfoQ by Bruno Couriol.
- QA Agent, Bug0 vendor page.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
