Sign InOpen Brain
AI EngineerVideoSource Linked

The Dark Arts of Web Automation: Teaching Agents to Use Websites Like Humans — Corey Gallon, Rexmore

Reliable browser agents pair a programmable CLI with Chrome DevTools Protocol input, then sense, act, and verify one move at a time. Save solved paths as deterministic code instead of re-reasoning.

AI Engineer · Aug 14, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The approach gives agents a CLI over Chrome DevTools Protocol, exposing the DOM, accessibility tree, screenshots, network traffic, logs, clicks, and keystrokes. Its core control loop is **sense, act, verify** across independent channels.

Practical Implication

Start with cheap programmatic actions, escalate to trusted browser input only when required, and use vision or human-like motion as the last rung. Once a path works, encode it so deterministic code handles repeated interaction while the model handles perception.

Agent-Ready Context
The approach gives agents a CLI over Chrome DevTools Protocol, exposing the DOM, accessibility tree, screenshots, network traffic, logs, clicks, and keystrokes. Its core control loop is **sense, act, verify** across independent channels.

Start with cheap programmatic actions, escalate to trusted browser input only when required, and use vision or human-like motion as the last rung. Once a path works, encode it so deterministic code handles repeated interaction while the model handles perception.

The demonstrations include bypassing **Turnstile, MTCaptcha, Lemon, and reCAPTCHA v2**, behavior that can trigger abuse controls and violate site policies. The method therefore needs explicit authorization and tight operational boundaries.
Connected Context · Feed7 Judgment

This makes browser automation a layered harness problem: combine independent observations, verify each action, and convert successful paths into deterministic routines. It complements visual-only feedback with DOM, accessibility, network, and log evidence, while its CAPTCHA demonstrations sharply increase the need for authorization, action boundaries, and approval controls.

Perception Agents — Antje Barth, Amazon AGI LabThe perception harness supplies one feedback channel; this Signal broadens verification across visual, DOM, accessibility, network, and log evidence.From RL to IRL — Gaurav Mishra, Amazon AGI LabIts sense-act-verify loop addresses ordinary interaction failures, while the candidate adds checkpoints, rollback, interruption, and handoff for persistent or risky consequences.Bringing agents onto the world wide web — Paul Klein IV, BrowserbaseBoth place browser reliability in the runtime harness; the candidate adds stable environments, authentication, site knowledge, and approvals needed around this interaction ladder.In Code They Act, In Proof We Trust — Erik Meijer, Leibniz LabsEncoding proven paths as deterministic code makes actions more inspectable, but the candidate imposes the stronger prerequisite of machine-checkable safety before side effects execute.
Context Map
agentcoding#computer-use#harness-engineering#tool-use
Uncertainty
The demonstrations include bypassing **Turnstile, MTCaptcha, Lemon, and reCAPTCHA v2**, behavior that can trigger abuse controls and violate site policies. The method therefore needs explicit authorization and tight operational boundaries.