All projects
Software & Agents
Browser Challenge Agent
A DOM-first Playwright agent that works through a 30-step browser challenge, with an optional LLM planner.
Personal project · Brett Adcock's 30-step browser challenge
Overview
A browser-automation agent built around one idea: understand the DOM before acting. Each step starts with a structured observation of the page, taken with page.evaluate(). Specialized solvers then handle the challenge types: reveal buttons, hidden DOM elements, delayed reveals, hover targets, math questions and modal choices.
An optional LLM planner (OpenAI) can be switched on, with token and cost reporting. The agent runs in a precise mode by default, with brute-force clicking and auto-progress switched off.
Highlights
- 01Rich, structured DOM observations instead of screenshots as the primary signal.
- 02Specialized solvers per challenge type, with an optional LLM planner.
- 03Per-run debug artifacts: an event log, a report with timing and cost, DOM snapshots and screenshots on stall.