Browser Harness is 592 lines of Python and the agent writes the rest
The browser-use team just dropped Browser Harness, a self-healing harness for browser tasks. 6.4k stars on GitHub as of today, 158 commits on main, still only about 592 lines of Python total. It runs straight on Chrome DevTools Protocol. No heavy framework on top.
The idea is deliberately thin. There's run.py (36 lines) that executes Python with helpers preloaded. There's helpers.py (195 lines) with the tool implementations. There's admin.py and daemon.py handling the CDP websocket plumbing. And there's a domain skills folder for site-specific playbooks like LinkedIn and Amazon. That's it. That's the whole stack.
The self-healing part is where this gets interesting. When the agent hits a task it can't do — a helper it needs that doesn't exist — it writes the missing function directly into helpers.py and keeps going. No crash, no operator-in-the-loop. The agent grows the toolkit as it encounters work. This flips the normal agent-framework model: instead of an elaborate library of pre-defined tools that the agent picks from, you ship a near-empty runtime and let the agent extend it per task.
This is the Karpathy take on agent design landing as shipped code. The strongest loop is not the one with the most primitives. It's the one that can grow its own primitives against a measurable target. For anyone building a browser agent today — and that is half the agent ecosystem right now — Browser Harness is worth reading top to bottom in one sitting. browser-use.com also gives you three concurrent cloud browsers free with no credit card if you want to try it without touching your own Chrome.
https://github.com/browser-use/browser-harness
← Back to all articles
The idea is deliberately thin. There's run.py (36 lines) that executes Python with helpers preloaded. There's helpers.py (195 lines) with the tool implementations. There's admin.py and daemon.py handling the CDP websocket plumbing. And there's a domain skills folder for site-specific playbooks like LinkedIn and Amazon. That's it. That's the whole stack.
The self-healing part is where this gets interesting. When the agent hits a task it can't do — a helper it needs that doesn't exist — it writes the missing function directly into helpers.py and keeps going. No crash, no operator-in-the-loop. The agent grows the toolkit as it encounters work. This flips the normal agent-framework model: instead of an elaborate library of pre-defined tools that the agent picks from, you ship a near-empty runtime and let the agent extend it per task.
This is the Karpathy take on agent design landing as shipped code. The strongest loop is not the one with the most primitives. It's the one that can grow its own primitives against a measurable target. For anyone building a browser agent today — and that is half the agent ecosystem right now — Browser Harness is worth reading top to bottom in one sitting. browser-use.com also gives you three concurrent cloud browsers free with no credit card if you want to try it without touching your own Chrome.
https://github.com/browser-use/browser-harness
Comments