I built a browser for AI agents, then benchmarked it against Playwright MCP and browser-use

Описание и применение

Создан браузер AgentGo для AI‑агентов, использующий компактные описания страниц и диффы, что сократило токены и время выполнения по сравнению с browser-use и Playwright MCP. В тестах MiniWoB++ AgentGo решил 22/24 задачи за 32 с и 94.7 к токенов.

Задача

AI agents reading too much of a page snapshot, wasting tokens on unchanged content

Результат

AgentGo solved 22/24 MiniWoB++ tasks, median 32 s, 94.7k input tokens, 7.7 turns, outperforming browser-use (19/24, 92 s, 209.9k tokens, 13.2 turns) and Playwright MCP (18/24, 42 s, 178.4k tokens, 13.0 turns)

Рабочий процесс

• 1. Agent opens page with browser_open, receiving a compact outline of headings and actionable elements
• 2. Agent performs actions via browser_act
• 3. Browser returns only the diff (+, -, ~) of changed elements
• 4. Browser waits for page to settle before responding, eliminating unnecessary wait actions

Доказательства

Every agent I tried on the web had the same problem: it was reading too much. A browser tool hands the model a page snapshot, the model clicks, and then gets a fresh snapshot of the whole page again.

Источник

Same model (gpt-6-luna, medium reasoning), same 24 BrowserGym MiniWoB++ tasks, a fresh agent session for every task, only that browser's MCP tools, and the page's own reward decides.

Источник

AgentGo 0.1.5 22/24 32 s 94.7k 7.7

Источник

Three design choices do most of the work: 1) A compact outline instead of a snapshot. 2) Only what changed. 3) The browser waits, not the model.

Источник