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

A brand new wave of AI-powered browser-use brokers is rising, promising to rework how enterprises work together with the online. These brokers can autonomously navigate web sites, retrieve data, and even full transactions – however early testing reveals important gaps between promise and efficiency.

Whereas shopper examples supplied by OpenAI’s new browser-use agent Operator, like ordering pizza or shopping for sport tickets, have grabbed headlines, the query is about the place the primary developer and enterprise use instances are. “The factor that we don’t know is what would be the killer app,” mentioned Sam Witteveen, co-founder of Pink Dragon, an organization that develops AI agent purposes. “My guess is it’s going to be issues that simply take time on the internet that you simply don’t truly get pleasure from.” This contains issues like going on the internet and trying to find the most affordable worth of a product or reserving the very best resort lodging. Extra possible it is going to be utilized in mixture with different instruments like Deep Analysis, the place firms can then do much more subtle analysis plus execution of duties across the net.

Firms must fastidiously consider the quickly evolving panorama as established gamers and startups take totally different approaches to fixing the autonomous shopping problem.

Key gamers within the browser-use agent panorama

The sphere has rapidly change into crowded with each main tech firms and modern startups:

Operator and Proxy are probably the most superior, when it comes to being consumer-friendly and out-of-the-box prepared. Most of the others look like positioning themselves extra for developer or enterprise utilization. For instance, Browser Use, a Y-Combinator startup that permits customers to customise the fashions used with the agent. This offers you extra management over how the agent works, together with utilizing a mannequin out of your native machine. However it’s undoubtedly extra concerned.

The others listed above present a various diploma of performance and interplay with native machine assets. I made a decision not even to check ByteDance’s UI-TARS for now, as a result of it requested decrease degree entry to my machine’s safety and privateness options (if I try it out, I’ll undoubtedly use a secondary pc). 

Testing reveals reasoning challenges

So the simplest to check are OpenAI’s Operator and Convergence’s Proxy. In our testing, the outcomes highlighted how reasoning capabilities can matter greater than uncooked automation options. Operator, specifically, was extra buggy.

For instance, I requested the brokers to search out and summarize VentureBeat’s 5 hottest tales. It was an ambiguous job, as a result of VentureBeat doesn’t have a “hottest” part per se. Operator struggled with this. It first fell into an infinite scrolling loop whereas trying to find ‘hottest’ tales, requiring guide intervention. In one other try, it discovered a three-year-old article titled “Prime 5 tales of the week.” In distinction, Proxy demonstrated higher reasoning by figuring out the 5 most seen tales on the homepage as a sensible proxy for reputation, and it gave correct summaries.

The excellence grew to become even clearer in real-world duties. I requested the brokers to ebook a reservation at a romantic restaurant for midday in Napa, California. Operator approached the duty linearly — discovering a romantic restaurant first, then checking availability at midday. When no tables had been out there, it reached a useless finish. Proxy confirmed extra subtle reasoning by beginning with OpenTable to search out eating places that had been each romantic and out there on the desired time. It even got here again with a barely higher rated restaurant.

Even seemingly easy duties revealed vital variations. When trying to find a “YubiKey 5C NFC worth” on Amazon, Proxy rapidly discovered the merchandise extra simply than Operator. 

OpenAI hasn’t divulged a lot about applied sciences it makes use of for coaching its Operator agent, apart from saying it has skilled its mannequin on browser-use duties. Convergence, nonetheless, has offered extra element: Its agent makes use of one thing referred to as Generative Tree Search to “leverage Net-World Fashions that predict the state of the online after a proposed motion has been taken. These are generated recursively to supply a tree of doable futures which can be searched over to pick out the subsequent optimum motion, as ranked by our worth fashions. Our Net-World fashions may also be used to coach brokers in hypothetical conditions with out producing numerous costly information.” (Extra here).

Benchmarks could also be ineffective for now

On paper, these instruments seem intently matched. Convergence’s Proxy achieves 88% on the WebVoyager benchmark, which evaluates net brokers throughout 643 real-world duties on 15 widespread web sites like Amazon and Reserving.com. OpenAI’s Operator scores 87%, whereas Browser-Use says it reaches 89% however solely after altering the WebVoyager codebase barely, it conceded, “in line with our wants”.

These benchmark scores ought to actually be taken with a grain of salt, although, as they are often gamed. The true take a look at is available in sensible utilization for real-world instances. It’s very early, the area is so quickly altering, and these merchandise are altering nearly every day. The outcomes will rely extra on the particular jobs you’re attempting to do, and you could need to as a substitute depend on the vibes you get whereas utilizing the totally different merchandise.

Enterprise implications

The implications for enterprise automation are important. As Witteveen factors out in our video podcast conversation about this, the place we do a deep dive into this browser-use pattern, many firms are at the moment paying for digital assistants – operated by actual folks – to deal with fundamental net analysis and information gathering duties. These browser-use brokers may dramatically change that equation.

“If AI takes this over,” Witteveen notes, “that’s going to be among the first low hanging fruit of individuals shedding their jobs. It’s going to point out up in a few of these sorts of issues.”

This might feed into the robotic course of automation (RPA) pattern, the place browser use is pulled in as simply one other device for firms to automate extra duties. And as talked about earlier, the extra highly effective makes use of instances will probably be when an agent mixed browser use with different instruments, together with issues like Deep Analysis, the place an LLM-driven agent makes use of a search device plus browser use to do extra subtle jobs.

Value dynamics driving innovation

One other key issue driving fast improvement is the provision of highly effective open-source reasoning fashions like DeepSeek-R1. This enables firms constructing these browser-use brokers to compete successfully with bigger gamers by leveraging these fashions somewhat than constructing their very own.

The pricing stress is already evident. Whereas OpenAI requires a $200 month-to-month ChatGPT Professional subscription to entry Operator, Convergence affords restricted free use (as much as 5 makes use of per day) and a $20/month limitless plan. This aggressive dynamic ought to speed up enterprise adoption, although clear use instances are nonetheless rising.

Safety and integration challenges

A number of hurdles stay earlier than widespread enterprise adoption. Some web sites actively block automated shopping, whereas others require CAPTCHA verification. Whereas OpenAI and Convergence have instruments that may get previous CAPTCHAs, they let customers take over the duty to fill them out — as a substitute of doing them immediately, for the reason that entire level of CAPTCHAs is to make sure a human is on the different finish. Instruments like ByteDance’s UI-TARS request deep system entry, which raises safety considerations for enterprise deployment.

Moreover, the method to web site cooperation varies. OpenAI has labored with particular companions like Instacart, Priceline, DoorDash and Etsy, whereas others try and navigate any web site. This inconsistency may influence reliability for enterprise use instances. And naturally, any time an agent hits a website requiring login particulars, that can gradual issues — because the brokers will flip issues over to you to fill in these particulars.

Wanting forward

For enterprises evaluating these instruments, the main target ought to be on particular use instances the place autonomous net interplay may present clear worth – whether or not in analysis, customer support, or course of automation. The expertise is progressing quickly, however success will rely upon matching capabilities to concrete enterprise wants.

As this area evolves, count on to see extra enterprise-focused options and probably specialised brokers for particular industries or duties. The race between established gamers and modern startups ought to drive each technical development and aggressive pricing, making 2025 an important yr for enterprise browser-use agent adoption.

For extra element on these traits and testing outcomes, take a look at the full video conversation between Sam Witteveen and myself.