Browserless Agent just killed the browser harness. One MCP tool. Six benchmarks won.

Six tools entered the browser-agent benchmark. Browserless Agent beat all six across every task.

18 seconds on TodoMVC. 2 minutes on complex form fills. 2 minutes 30 seconds on Amazon multi-item add — where three competitors failed entirely.

This is not another browser automation wrapper.

The architecture is the breakthrough. Most MCP browser tools expose 60+ discrete tools. Browserless Agent is one tool. One stateful session. 25+ commands via { method, params }.

The session stays open. Cookies survive. Login state persists. Page context carries forward across every tool call.

Command batching collapses multi-field forms from 10+ round trips into 3. That is not a marginal improvement — it is the difference between a 2-minute form fill and an 8-minute one.

Here is what the other tools look like:

Playwright MCP: 60+ tools, no batching, per-profile stateful. Browser Use: Python harness, vision-based, stateless HTTP. Stagehand: Node SDK required, mixed page representation. ChatGPT native: black box, no batching, no stealth.

Browserless Agent: one config block in Claude Desktop, Cursor, VS Code, or Windsurf. Zero harness code.

The accessibility-tree snapshot is the unsolved problem they cracked. CSS selectors break when pages shift. Raw DOM is too large for LLM reasoning. The accessibility tree is compact, stable across redesigns, and built on ARIA roles that browsers cannot change without breaking their own compliance.

Stealth rendering, CAPTCHA solving, and residential proxy routing are handled at the infrastructure layer — not as add-ons, not as SDK patches. Built in.

If you are building agents that need to navigate real consumer apps — Google Flights, Amazon, finance portals — this is the baseline now. Anything slower than 2 minutes on a multi-step form is a bug.

Audit your browser automation stack. If you are running 60+ Playwright tools or stitching together a Python harness, you are burning tokens on tool selection overhead that could be spent on reasoning.

SOURCE: https://www.browserless.io/blog/introducing-browserless-agent
VERIFIED: browserless.io/blog (primary source), browserless.io/changelog/2026-06, techstackups.com browser benchmarks
SIGNAL: The browser-agent interface just consolidated from 60+ tools to one. This changes the economics of every agent that needs to touch the live web.