Agent Reading Test: New Benchmark Reveals How Poorly AI Coding Agents Read Web Documentation

Hacker News·medium signal

A new benchmark at agentreadingtest.com tests how well AI coding agents (Claude Code, Cursor, Copilot, etc.) actually read web documentation, surfacing silent failure modes: content truncation, CSS-buried text, client-side rendering delivering empty shells, and tabbed content serializing into walls of text. The test gives agents 10 realistic documentation tasks with embedded canary tokens — a perfect score is unlikely for any current agent. Companion project to the Agent-Friendly Documentation Spec with 22 checks across 8 categories.