DEMO / RECORDING
Overview & Result
Dan Leshem reports rebuilding Stencil's QA agent around Jev: a frontier model plans the test, Jev chooses browser actions, and vision checks the result. The shown run signs in, opens an app, triggers a paywall, checks a spacing fix, and passes five checkpoints in 28.3 seconds with a screenshot and trace for each step.
How Jev fits in the loop
- The frontier model plans the QA scenario.
- Jev selects the next bounded browser action.
- Vision checks the rendered result while the runtime decides whether to continue, retry, prioritize, or request human review.
How to reproduce
- Open the original Stencil QA post and inspect the attached video as the primary evidence.
- Model the loop as frontier-model planning, Jev browser-action choice, and vision-based verification.
- Treat the reported timing as a target to remeasure; do not present it as an independent benchmark.
Why this build matters
This separates repeated action/evaluation decisions from generation and visual inspection inside a browser QA loop.
Limitations
- The post does not link a public repository or a runnable reproduction.
- The 28.3-second result is an author-reported run, not a JevForAgents measurement.