The AI browser agent Jev Ultrafast chooses to read code (DOM) directly instead of analyzing the entire screen as an image, reducing costs and increasing speed by over 25%.
Imagine this: On a busy morning, you tell your AI assistant, “Find and book the cheapest flight to London for next week,” and go off to get coffee. The AI instantly traverses numerous airline websites, finds the cheapest ticket, and completes the payment. In the past, this sounded like something out of a sci-fi movie, but now, AI browser agents are starting to take on that role. However, there has been one major problem: the way AI “sees” websites has been too slow and inefficient.
The recently emerged Jev Ultrafast (Ref 1) is a new browser agent stepping up to solve exactly this problem. Today at MindTickleBytes, we will easily explain why this technology is important and how it will change the way we use the web.
Why is this important?
Many existing autonomous web agents have relied on “vision” to understand websites, just like humans. They repeatedly captured the screen every moment and asked the AI, “What do you see on the screen now?” before waiting for an answer. This is just as inefficient as us taking a photo of a smartphone screen every second to analyze it.
Jev Ultrafast has boldly abandoned this “image capture-analysis” loop (Ref 2). This is more than just a technical improvement; it increases the speed at which AI uses web services by over 25% (Ref 2). Not only is the wait time for users shorter, but the computing costs required to run the AI are drastically reduced, laying the foundation for AI assistants to get much closer to our daily lives (Ref 6).
Simplifying: It reads the ‘blueprint,’ not the ‘image’
To use a simple analogy: if an existing agent were like someone checking for a building by taking photos of its exterior one by one, Jev Ultrafast is like someone holding the building’s “blueprint” directly in their hand.
Websites are ultimately composed of complex code that computers can read, called DOM (Document Object Model). Jev Ultrafast extracts the structure of this code as a “snapshot” and organizes the elements within it neatly into an index (number) (Ref 2).
In short, instead of “showing” the website to the AI every time, it proposes, “I’ll give you the layout of this website, so choose the button number from here.” To achieve this, it uses TypeSafe’s ‘Jev Choice’ technology, allowing the AI to perform tasks smoothly without stopping to ponder once a goal is set (Ref 9).
Of course, in special situations where the AI needs to input text (e.g., entering a date in a search bar), a small language model is brought back in to handle things flexibly (Ref 1, Ref 6). It is equipped with a smart division-of-labor system that utilizes appropriate tools depending on the situation.
Current status: How far has it come?
Jev Ultrafast is already proving its real-world performance. In an actual demonstration, it completed the task of searching for a flight from Zurich to London using Google Flights in just 7.1 seconds (Ref 1, Ref 6). The cost incurred for this process was approximately $0.0039—an amazing level of efficiency (Ref 6).
Jev is often referred to as a “System One Model,” meaning it is optimized for making immediate decisions without complex thought, much like the system in the human brain that reacts quickly and unconsciously (Ref 5). However, there is a caveat. As with any technology, in the early stages, it is sometimes reported that website structures change unexpectedly or tasks stop (BLOCKED state) when data is not returned properly while using libraries (Ref 8). In other words, we must remember that it is a promising technology that has only just started to take its first steps.
What’s next?
In the future, AI agents will go beyond simply searching for information on our behalf; they will handle complex web-based tasks like shopping, reservations, and management more quickly and cheaply. The pace of technological development is very steep, with some claiming it is 200 times faster (Ref 14).
Someday, you will leave a short sentence like “Prepare for my next vacation,” and find a screen where the AI agent has instantly finished everything from flight reservations to hotel confirmations. What we need to pay attention to now is not just how much “smarter” AI becomes, but how much more “efficiently” it uses our tools, just like this.
MindTickleBytes’ AI Reporter View
Jev Ultrafast presents an important turning point in how AI handles human tools. The shift from existing methods that relied solely on visual perception to utilizing structural data will be a practical bridge helping AI agents quickly integrate into real-world tasks.
References
- GitHub - browser-use/jev-ultrafast (https://github.com/browser-use/jev-ultrafast)
- Jev Ultrafast Cuts Browser Agent Time by 25% With TypeSafe … (https://news.lavx.hu/article/jev-ultrafast-cuts-browser-agent-time-by-25-with-typesafe-action-space)
- Jev Ultrafast: A browser agent with a dynamic, indexed action … (https://news.ycombinator.com/item?id=49735979)
- How Does Jev Work? RLCD & Parallel Inference Explained … (https://www.explainx.ai/blog/how-does-jev-work-rlcd-system-one-model-explained-2026)
- [AINews] Jev: a “System One Model” that only decides … (https://www.latent.space/p/ainews-jev-a-system-one-model-that)
- Gregor Zunic on X: “Breaking: Browser Use + Jev = Ultrafast ⚡ … (https://x.com/gregpr07/status/2100411066966749359)
- browser-use/jev-ultrafast — GitHub trending stats & insights (https://trendshift.io/repositories/242003)
- Library API: first observation can return an empty action space; agent terminates with BLOCKED instead of retrying (https://github.com/browser-use/jev-ultrafast/issues/1)
-
vlad-terin/jev-browser DeepWiki (https://deepwiki.com/vlad-terin/jev-browser) -
jev-browser-mcp by Ying-Kai-Liao Glama (https://glama.ai/mcp/servers/Ying-Kai-Liao/jev-browser) - Building Browser Agents: Architecture, Security, and Practical Solutions (https://arxiv.org/html/2511.19477v1)
- BrowserAgent: Building Web Agents with Human-Inspired Web Browsing Actions (https://arxiv.org/html/2510.10666v2)
- Best 30+ Open Source Web Agents (https://aimultiple.com/open-source-web-agents)
- Jev: TypeSafe’s Decision Model, Speed and Cost Explained (https://www.orcarouter.ai/blog/jev-typesafe-system-one-what-we-know)
- It captures the screen as an image every moment
- It reads code (DOM) directly to structure it
- It records human clicks directly
- Because it is a text-generation-centric model
- Because it is a very fast image-processing model
- Because it makes rapid decisions in a non-autoregressive manner
- 7.1 seconds
- 25 seconds
- Over 1 minute