Submit a projectWe compared two ways of selecting browser controls: a Codex call at each step and a Jev selection call. Both used the same real mouse/keyboard executor and page-state checks in an isolated Chrome window.
What we tested
Three local, synthetic workflows: settings to network to DNS; finding a report for a given period; and entering a search query. Each had three steps and ran twice per configuration. Run order alternated and paired layouts used the same seed.
| Measure | Codex per step | Jev assisted |
|---|---|---|
| Completed workflows | 6/6 | 6/6 |
| Correct selections | 18/18 | 18/18 |
| Total after page ready | 88.607 s | 16.099 s |
| Including page navigation | 91.677 s | 19.160 s |
What the timing does not prove
The Codex configuration started a fresh CLI process for each decision and carried a full agent context. Jev used direct HTTPS. The pages were simple and each immediate subgoal was supplied. This tests control selection within a prepared workflow, not autonomous planning on arbitrary sites or visual grounding.
The navigation-inclusive time ratio was about 4.78 in this configuration. Real network waits, harder interfaces and different calling overhead will change it. A deterministic script can also complete a known path quickly.
A useful design pattern
Keep planning and verification with the main agent. Try a smaller decision layer for frequent, bounded selections. Re-check that the chosen control still exists and is enabled, then verify the actual result after execution. Return uncertain cases to the main agent.
Tested September 18, 2026 with Jev 1.13.0, GPT-6 Astra (low reasoning), Chrome 153 and Playwright 1.58. This small sample does not establish a general speed or reliability improvement.
Download the sanitized timing data (CSV)