The rise of browser-use agents: Why Convergence’s Proxy is beating OpenAI’s Operator

A new wave of agents used an AI-powered browser that promises to transform the Internet on the Internet. These agents are able to visit websites, access and even get full operations – but early test promises and can detect significant gaps between performance.

Consumer patterns using Openai’s new browser, like to order pizza or buy game tickets, the main developer and the enterprise are about where the business is working. “What we do not know, there will be a murder application – said Sam Witteveen, a company developed by the AI ​​Agent, the co-founder Sam Witteveen said. It includes things like searching for the cheapest price or ordering the best hotel habitats. It will be used in conjunction with more other Tools such as deep researchThen companies can make a further advanced investigation plus Execution of tasks around the web.

Companies should carefully appreciate the rapidly developing landscape such as players and beginnings and take different approaches to the solution of the autonomous crawl.

The main players of the browser – agent landscape

The area is quickly crowded with both large technological companies and innovative beginnings:

The most advanced operator and proxy, in terms of consumer friend and ready beyond the box. Many of the others places themselves more for many developers or enterprise use. For example, The use of browserA combination start that allows users to make models used with agent. This allows you to manage how it works, including the use of a model from the agent’s local machine. But it is definitely more involved.

Others listed above have different functionality and interaction with local machine resources. Due to the safety and lower access to the protection of the machine, this is to test the UI-TARS of the Privacy, because I have a lower access to the security and privacy properties (if I try it, I will use the secondary computer ).

The test reveals grounding problems

Thus, the easiest is the easiest Openai operator and proxy of approximation. The results of the results of the raw automation of the raw automation of the raw automation in our testing are stressed. The operator was especially occasionally occasionally.

For example, I wanted to find and summarize the most popular story of VentureBeat from agents. This was an indefinite task because Venturebeat has no “most popular” section per. The operator struggled with this. At first, when searching for “the most popular” stories, he fell into an endless slip loop while searching for “the most popular” stories. In another attempt, a three-year-old article was found “Top five stories of the week“On the contrary, Proxy demonstrated the most visible story in the home page with a practical reason for popularity and a summary.

The difference was clearer in real world assignments. Agents asked Napa to order reservations in a romantic restaurant in California. The operator approached the position as a line – to find a romantic restaurant first, then check the availability of the afternoon. When there is no table, it ended dead. Proxy showed more complex thinking starting with a rogue to find both romantic and any time available at any time. Even returned with a slightly better appreciated restaurant.

Even simple-looking tasks also revealed important differences. While searching for “Jubikey 5C NFC price” in Amazon, Proxy quickly found the item easier than the operator.

Openai has not been published in addition to the technology used by the operator to prepare the operator agent, otherwise taught the browser’s tasks. At the same time, the rapprochement was reported: Agenti, after the proposed measure, uses something that uses something according to web world models that predicted the status of the Internet. These are recursively created to produce the possible futures tree, which is possible to choose the next optimal movement, which is ranked by our value models. Our web world models can also be used to train agents in hypothetical situations without creating very expensive information. “(More here).

Assessments can now be useless

These tools in the paper seem to be closely fit. Convergity of Convergity Reaching 88% on Webvoyager BenchmarkEvaluating web agents between 643 real world tasks on 15 popular sites like Amazon and Booking.com, evaluate web agents. Openai operator 87%, while using the browser It says you will reach 89% But only after Webvoyager Codebase is slightly changed “for our needs.”

Although these criteria points should really be taken with a grain of salt, they can also come. The real test is in practical use for real world cases. Very early, the space changes so fast, and these products change almost daily. The results will be more dependent on special work you are trying to do, and instead you may want to trust the vibes you receive when using different products.

Enterprise results

The effects for enterprise automation are significant. As you sign in Witteveen Video Podcast conversation Many companies that we will use this browser are currently managed for virtual assistants – are managed by real people – to manage the main web research and collecting tasks. The agents of using this browser were able to change this equation sharply.

“If the AI ​​takes it,” Witteveen Notes “will be the first low-hanging fruit from people who lost their jobs. It will appear in some of such things.”

This can be fed to the automation of the robot process (RPA) trend, which is used as another tool for the use of companies to automate the task. As noted earlier, an agent will have more powerful use when used by other means of combined browser, including other means Deep researchWhere a LLM-in-the Agent search vehicle uses plus Use the browser to do more complicated work.

Dynamics Management Update

Another key factor managing rapid development is the availability of powerful open source source such as DeepSeek-R1. This allows companies that use this browser to effectively compete with larger players using these models.

Price pressure is already clear. Openai provides an unlimited free use of Convergence, when required a Chatgpt Pro subscription to the Access operator $ 200 monthly. This competitive dynamics must accelerate the admission of the enterprise, although clear use is still being created.

Problems of security and integration

The widespread enterprise remains several obstacles before adoption. Some websites require automated crawls, others also check the CAPTCHA. Although Openai and Convergencence had the tools that the former CAPTCHA could get the users, they allowed us to complete them – instead of doing it directly, because the whole point of the captcha is that the whole point of CAPTCHA is in the other end of a person. Byje byTettEtace’s UI-TARS requires a deep system access to the fact that the ventures increase security concerns for enterprise placement.

In addition, the approach to the website changes. Openai has Worked with special partners such as InstaCart, Priceline, Doordash and EtsyOthers are trying to walk on any website. This discrepancy may affect the reliability for the use of enterprise. Of course, you always hit a site that caused an agent’s entry detail, and the agents will turn these details to you.

Looking forward

For the evaluating instruments, the focus must be in particular use of autonomous web interactions that focus on the focus – research, customer service or process automation. Technology is rapidly progressing, but success will depend on opportunities to suit the needs of specific businesses.

As this space develops, expect more enterprise-oriented properties and potential specialized agents for special industry or tasks. The race between the built-in players and innovative beginnings must manage both technical progress and competitive prices, preparing a very important year for the use of 2025 enterprise browser.

For more information on these trends and test results, check, check Sam Witteveen and full video chat between myself.