How to Build a Fast Browser Agent with Jev

TL;DR

Combine Jev’s rapid classification with Claude’s planning, writing, reasoning, and verification to automate browser tasks efficiently. Jev selects actions and numbered page elements from structured browser data, while executable scripts perform the actions. Low-confidence, blocked, text-entry, and completion states are escalated to Claude, creating a fast loop that remains independently checked.

Transcript

Browser agents have always been terrible. A 2-minute job for you as a person often takes 10 to 20 minutes using them, which means you may as well do the task yourself. But with Jev, when you set it up correctly, it's totally different. Every click can be decided in about a third of a second. And there's a multitude of brilliant examples showing it ... Read More

Key Insights

• Jev is a judgment model that answers predefined questions about an input state. It returns an option, probability information, and confidence rather than conversational text, allowing it to make thousands of small classification decisions quickly and at low input cost.
• Jev supports three question types: yes-or-no questions for deciding whether to act, choices for routing or selecting among defined options, and descriptive scales for scoring and sorting data. The required use of the answer determines which question type is appropriate.
• Fast browser navigation depends on structured page data rather than screenshots. Playwright or Jev Ultrafast extracts interactive elements, numbers them, and sends their descriptions to Jev, avoiding the time required to load pixels and reason visually about every page.
• Browser actions are selected through two central choice questions. One identifies the action type, such as click, type, select, scroll, wait, done, or blocked. The other identifies the numbered page element, with an option for cases where no suitable element exists.
• Confidence thresholds control whether an action is executed automatically. If both the action and element answers exceed a configured threshold, such as 0.7 in the example, deterministic scripts perform the action, wait briefly for the page to settle, and collect the next state.
• Claude and Jev perform complementary roles rather than replacing one another. Jev handles frequent, constrained decisions, while Claude plans workflows, reasons through exceptions, produces required text, adjusts questions, validates responses, and independently checks whether the requested browser task is actually finished.
• Executable scripts make accepted browser actions deterministic. The system defines these scripts during planning, so clicking, scrolling, selecting, and other routine operations run as code after Jev chooses the appropriate action and target, without requiring another artificial intelligence decision.
• The combined approach substantially reduced browser calls in the cited Google Flights run. Browseruse.com reportedly decreased the number of calls from 1,092 to 101 by combining a cheap, fast language model with Jev for deterministic decisions on web pages.

Questions & Answers

Q: How does Jev make browser agents faster?

Jev makes browser agents faster by evaluating structured descriptions of interactive page elements instead of processing screenshots or generating conversational responses. Browser tooling numbers elements such as links, buttons, and input boxes, then Jev selects an action and target from predefined options. Deterministic scripts execute high-confidence decisions, after which the updated page is collected and evaluated again.

Q: What question types can Jev answer?

Jev can answer three kinds of questions. A yes-or-no question determines whether an action should occur. A choice question selects one option from a supplied list and can route a query or choose a browser action. A scale question assigns a score described in words, which is useful when information must be sorted by the resulting output.

Q: How do Jev and Claude work together in browser automation?

Claude creates the plan, handles reasoning, writes text when needed, and verifies results, while Jev performs many small and rapid classification decisions. Jev can choose an action and page element, but it cannot generate the contents of a text field. Claude also intervenes when confidence is insufficient, progress is blocked, or completion needs independent confirmation.

Q: How does Jev choose what to do on a web page?

The system sends Jev structured page data and two main choice questions. The first asks which action to take, with options such as click, type, select, scroll, wait, done, or blocked. The second asks which numbered element should receive that action. Jev returns answers with confidence values that determine whether executable scripts may proceed.

Q: What happens when Jev has low confidence?

When either returned confidence falls below the configured acceptance level, the automated loop stops instead of executing the uncertain action. Claude then evaluates the original goal and the progress made so far. It can revise the questions sent to Jev or use its own reasoning to help the workflow recover from the uncertain or blocked situation.

Q: Why does the browser agent use numbered page elements?

Numbered page elements turn a dynamic interface into structured input that Jev can classify quickly. Playwright or Jev Ultrafast records interactive items such as pricing links, trial buttons, and email fields, then assigns each an identifier. Jev can select one of these identifiers without examining screenshots, loading pixels, or reasoning visually about the entire browser page.

Q: How does the system verify that a browser task is complete?

The runner does not automatically trust Jev when it selects the done action. Instead, the state is passed back to Claude for an independent check against the original goal. This verification step is important because Jev only makes constrained judgments, while Claude can assess the broader task history and determine whether the requested outcome was actually achieved.

Q: What is needed to build a Jev browser agent?

The demonstrated setup requires a Type-safe key in the environment file, a Jev or Type-safe skill that teaches Claude how to phrase and batch questions, and a prompt or implementation plan. Claude Code can then build a workflow that combines Claude, Jev, browser control, executable scripts, screenshots, and flow documentation for selected websites.

Summary & Key Takeaways

• Jev accelerates browser automation by classifying structured page elements instead of interpreting screenshots or generating text. Playwright or Jev Ultrafast collects and numbers interactive elements, then Jev selects an action type and target. This approach lets the agent make page-level decisions in a fraction of a second.

• The automation loop begins with Claude planning the task, followed by browser tooling collecting page data. Jev answers batched questions about the appropriate action and element, and deterministic scripts execute accepted decisions. Each page change creates a new state, allowing the same classification and execution cycle to continue.

• Claude handles situations that require broader reasoning or language generation. It intervenes when Jev reports low confidence, encounters a blocked state, needs text for an input, or claims the task is complete. Independent completion verification prevents the system from accepting Jev’s definition of done without checking the original goal.