Journal
Journal entry
Bridging Test Plans at Firefox with a Small, Boring Pipeline
How we wired an agentic GPT-4 workflow into Firefox QA without rewriting the workflow. The boring parts mattered more than the model.
Filed underMozillaAgentic AIQA
How we wired an agentic GPT-4 workflow into Firefox QA without rewriting the workflow. The boring parts mattered more than the model.
The most useful thing we did on the Firefox QA tool wasn't picking the right model. It was deciding to never replace a test plan, only augment one.
The pipeline is small: feature description in, a handful of seed cases in, candidate cases out, cosine-similarity diff against the existing plan, and a triage view that lets a human accept, edit, or discard. That last step is the product.
Models change. The triage view is the moat.