Use case
Jev for Browser Agents
How a browser agent can keep Jev inside a bounded choice: pick an operation and a control the page already listed.
Last verified: 2026-09-24
Overview
A browser agent has to decide the next click, type, or scroll without inventing selectors, coordinates, or scripts. Latency matters, and a wrong free-form action is hard to check.
Why Jev fits
- The action space is bounded: the harness lists controls, and Jev chooses among them.
- The loop is latency sensitive. One Choice request can name both the operation and the target.
- The output is structured, so the executor can reject anything that was not in the snapshot.
- Confidence on Choice can sit next to the decision, but a DONE answer is still not proof the task succeeded.
Real projects
Example workflow
- Read the visible controls into a numbered table.
- Ask Jev which operation applies, and which listed element it applies to.
- If the operation is TYPE_TEXT, ask a separate text model for the string.
- Execute only after the harness rechecks the page, then verify the outcome outside the model.
When Jev works well
- Tasks whose next step is a choice among controls the program can already see.
- Flows where a text model should write words, and Jev should only decide that typing is the next step.
When Jev may not fit
- Pages that depend on shadow roots, iframes, canvas, uploads, or pop-up tabs the current reader does not cover.
- Tasks that need the model to invent a plan, write code, or explain its reasoning.
Related tutorials
Browser-agent tutorialUltrafast architectureDecision layerChoice
Sources
- jev-ultrafast READMEBrowser Use · accessed 2026-09-24 · github
- TypeSafe speculative fan-outTypeSafe · accessed 2026-09-24 · documentation