Image: Bug0
Why it mattersKeeping a browser test suite stable is where end-to-end testing usually dies. The design pays a model price once per action, then runs cached actions straight through Playwright on every repeat. Assertions still run two models every time.
Bug0 open-sourced the engine behind its paid QA service. The project is called the Bug0 QA Agent. It is a Playwright library a developer imports inside an existing test suite. Tests are written as plain-English steps. The agent drives the browser on the first run and caches each single-action step. Later runs replay the cached actions straight through Playwright, with no model in the loop for those steps.
Bug0 says the library is already running in CI for its own paid customers. The company says 200-plus engineering teams use the service built on top of this core. Bug0 is backed by Accel, by Guillermo Rauch of Vercel, and by Naval Ravikant. The team was founded by Fazle Rahman and Sandeep Panda. Both were earlier at Hashnode, the developer blogging platform.
What runs on the first pass and what caches
A test is a list of named steps and assertions. A single-action step reads like "Click Acme Circles T-Shirt" or "Add to cart". On the first run an AI agent reads the page's accessibility snapshot and screenshots. It drives Playwright to carry out each step. Each single-action step is cached to Redis once it lands cleanly.
Later runs replay cached actions as direct Playwright calls. Bug0's own figure for first-run cost is about 30 seconds per action step. Cached runs replay in milliseconds. When the UI moves and a cached action breaks, the agent re-discovers the right interaction and updates the cache. Multi-action steps are not cached. They run through the agent every time.
Assertions run two models each time
Assertions never cache. Every assertion is checked on each run. Bug0 runs the check against Claude and Gemini in parallel. A third model breaks ties. The library reports a confidence score from 0 to 100 and passes only on consensus. A video flag records the whole run and sends the file to Gemini. The company says this catches a toast or an error message a single screenshot would miss.
The project is on GitHub under the name bug0inc/qa-agent. Bug0 calls the licence source-available. The library needs Node.js 18 or later, Playwright 1.59 or later, and API keys for Anthropic and Google AI. Redis is optional. Without Redis, action caching and cross-test state turn off and the rest works.
Bug0 is honest about one trade. A team saves a model call on cached action steps but still pays the model price on every assertion. A busy suite with many assertions is still a running cost. The win is where end-to-end suites usually die. A selector-based test breaks the moment a developer renames a class. A flake nobody traces gets muted. A suite that stops gating anything becomes decoration. The agent re-finds the button when the UI moves it.
Source
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.
