founder @browser_use
As models get better, the browser harness has to change. A lot. When we launched Browser Use in November 2024, GPT-4o wasn’t trained for computer use. Built on next-token prediction, it didn’t
A benchmark is a dataset and a judge. The judge sounds like a detail. It is the hard part. Everyone treats the judge as an afterthought, but it decides every number you publish, and a weak one makes
Your company is not a family. Families are unconditional. That is exactly how families should work. Companies are closer to championship sports teams. Players care deeply about each other, probably
How hard can it be to build the biggest company in the world? Insanely hard. But somebody does it. That is the whole point. When you decide something is impossible you are ngmi. You cannot win if you
I was reading about the early days of PayPal and had one thought: My God, these people were so delusional. Elon thought X.com would replace banks, insurance, basically everything. To
Running a startup is hard because there are thousands of things you could be working on at any moment. One minute you are coding, then a customer messages you, then there is a hiring problem, then
Don't wrap the LLM. Don't wrap its tools either. All you need is a SKILL.md and some Python helpers. The LLM has complete freedom. If something's missing, it writes it. The learning A few months ago
Every agent that visits a website starts from scratch. We made them remember. You visit Google Flights, type a city, hit enter - nothing happens. You have to wait for the dropdown, then click the
Training beats prompting. Every time. We gave our agent a Python tool and a browser. The browser crashed. The agent inspected the runtime object, found the dead browser instance, and wrote Python code
This is deep dive into our highly scalable LLM-as-a-judge evaluation platform that solves web agent variance. Evaluating an agent that browses the real web isn't like evaluating an LLM on text
All the value is in the RL'd model, not your 10,000 lines of abstractions. An agent is just a for-loop of messages. The only state an agent should have is: keep going until the model stops calling
