AI is building me a web browser
Claude built SWB, a 400,000-line browser with its own JavaScript engine, in seven weeks without its creator reading a single line of code.
Claude built SWB, a 400,000-line browser with its own JavaScript engine, in seven weeks without its creator reading a single line of code.
One thoughtful comment unlocks the ability to create threads.
Cool experiment, though I don't think it proves what it claims? The author still provided the "harness" (target sites, rules, and Chromium to verify it, not really 'fully autonomous'). It just proves the harness still does the heavy lifting.
My 2 cents: a sloppy harness vs. a solid one (clear specs, task breakdown, review checks) gives different level of control on results on the same budget. Probably this is where engineers still earn their keep right now, like ensuring the predictability of product quality, potentially reduce model cost, by applying the software engineering methodologies. Though as models get better and cheaper, who knows how long that'll last.
i guess with single agent and 500K lines of code he is going stall or is burning credits much-much quicker not going especially anywhere
lots of people work on scaling, from super complex to a simple mailbox/bus (claude code for example), he'd be much better even with that, many claude sessions running under one "PM" with a good design and plan before coding by the same team of claudes, maybe he even didn't need all these 500K lines of code after all :)
and i'm not against vibe coding, that is what I do too , but to send a prompt (no matter how detailed) to agent and hope it'll deliver something not mediocre is probably too optimistic
I’m a huge fan of the new Claude Code projects feature PMing a project for me, this entire forum is just in one long running chat that delegates building to various subagents!
yep, me too. the only difference i started doing that when nothing was available, so i built my own protocol for independent multi-agent communication. the difference is I use many agents from pretty any US vendor, todays team is claude (2-3) codex, gemini. same idea as project - agents in their own sessions do what "PM" assigns based on their capabilities and each agent has a decent idea of entire project
i've being doing that for almost a year, love it. i think doing that "human organization" like teams, agent selection on needed for task abilities, peer reviews from other agents, tests run by other agents, etc, makes many things so much simpler and results significantly better
protocol is agentchute on github :)
It's insane that building a browser has become the new building a todo list. Cursor had a good write up of their version of this from January (using GPT-5.2!): https://cursor.com/blog/scaling-agents
A dev built a full mostly working web browser without using up his weekly limits on a $100/month Claude Max plan. 400k lines of Rust and 500k lines of tests. He banned subagents because they kept trying to edit the same parts of the codebase leading to conflicts and burning tokens, so he only allowed one agent at a time and finished in two months.