Why It Didn't Work: Browser Agents

Bailey and I spent a little more than a month building better browser agents before pivoting away. Our initial thesis was:

• clearly people wanted browser agents: Manus was very popular, this was a hot area of research at the labs
• we had simple ways to improve on existing techniques (eg giving it better pre-processing of the DOM)

Tech Searching for a Problem #

We made a canonical startup mistake: decide what tech you want to build, then figure out who will use it. In our defense, we had a first customer who trialed our product from almost day one, but ACV would’ve been four digits, and there weren’t obviously thousands of customers with their same problem.

People Wanted Illegal Things #

When I did get around to doing more sales, people wanted to use our agents to access websites that either had a paid API (X, Epic) or very intentionally had no API (LinkedIn). We were on track to get sued the second we found success.

No Defensibility #

Several customers told us our browser agents were much better than what others offered out of the box at the time. This was entirely a result of using cute tricks like parsing the accessibility tree well. It took Anthropic ~4x longer to catch up than I expected. But that was still 10x too little time to build a durable business.