# Autonomy Bench

> Frozen single-HTML browser-autonomy prompts for coding agents. Humans use the gallery UI.

Suite: browser-autonomy-v2 2.4.0
Site: https://benchmark.gvaste.dev
Machine pack: https://benchmark.gvaste.dev/agent.json
WebMCP: the live gallery registers tools on document.modelContext (list-benches, get-bench-prompt, list-takes, show-take, get-gallery-state, get-receipt).

## Rules

- Use the exact prompt text for the requested level. Do not expand, clarify, or add hidden requirements.
- One cell is one benchmark, one prompt level, one model, and one attempt.
- Do not read sibling takes or published gallery HTML while producing a take.
- Return one self-contained HTML file named index.html. No build step.
- Unknown provenance stays the string not captured. Do not invent timings or token counts.

## Live benches

### Rollercoaster (`rollercoaster`)

#### A — Raw

Create a first-person rollercoaster with continuous playback in a single HTML file using Three.js.

#### B — Autonomous

Create a first-person rollercoaster with continuous playback in a single HTML file using Three.js. Make design, pacing, composition, and implementation decisions yourself wherever details are unspecified, turning the idea into a coherent finished experience.

#### C — Showcase

Create a first-person rollercoaster with continuous playback in a single HTML file using Three.js. Make design, pacing, composition, and implementation decisions yourself wherever details are unspecified, turning the idea into a coherent finished experience. Push toward presentation-ready quality with stronger art direction, motion, depth, transitions, atmosphere, and meaningful detail while preserving clarity and coherence.