browser-ai

# browser-ai

A browser agent in pure Go (stdlib only) driven by **Jev**, TypeSafe's System One model.
Jev does not generate text — it picks one option from a finite, indexed action space, so a
decision is one cheap round-trip. A small OpenAI-compatible LLM writes field values only when
the chosen operation is `TYPE_TEXT`.

Chrome is driven over CDP with a hand-rolled client: **launched over `--remote-debugging-pipe`**
(no TCP, no framing) or **attached to your real running Chrome** over a minimal WebSocket
client. Everything is instrumented so you can see where each step's time goes.

Reference implementation: [browser-use/jev-ultrafast](https://github.com/browser-use/jev-ultrafast)
(Python, MIT). The DOM snapshot script and the policy instructions are ported from it.

## Performance

Real runs, headless Chrome 153, this repo, 2026-09-20. Every pass is verified from the final page,
not from the model saying DONE. Full tables and method in [docs/BENCH.md](docs/BENCH.md).

Three runs per model, same goal text, pass = verified from the final page.

| task | decision maker | pass | median | per decision |
|---|---|---|---|---|
| **Google Flights** ZRH→LON one-way, until results visible (10 steps) | **Jev** | **3/3** | **9.6 s** | **330 ms** |
| | gemini-2.5-flash | 3/3 | 23.0 s | 1.06–1.16 s |
| | gpt-5.4-mini, reasoning low | 2/3 | 29.4 s | 1.3–1.7 s |
| | gpt-5.4-mini, reasoning none | 0/3 — one run had everything set but declared DONE before Search | 22.0 s | 0.85–0.91 s |
| | gemini-2.5-flash-lite | **0/3** — same wrong path every time, then a false DONE | 13.8 s | 0.9–2.3 s |
| | gpt-5.4-nano (none / low) | 0/3 · 0/3 | 16.7 s · 37.4 s | 0.8–1.7 s |
| **Wikipedia** → Gödel's incompleteness theorems (2 steps) | **Jev** | **3/3** | **3.4 s** | **330 ms** |
| | gemini-2.5