Show HN: An AI agent runs ten web errands a day, every transcript published

Agentability runs a daily public experiment that turns morning search trends into ten real “errands” and sends an AI agent to complete them using only plain web requests. The agent is strictly read-only (HTTP GET), with no JavaScript, logins, forms, retries, editing, or human assistance; every transcript is published verbatim so successes and failures are visible. Episodes report how many errands finished, pages read, sites visited, and where the agent hit “bot walls” or gave up, illustrating concrete failure modes when an assistant tries to find simple facts like kick-off times, quake magnitudes, or ticket prices.

Alongside the show is a weekly audited index of 113 well-known sites scored for agent readiness against checks agents rely on (llms.txt, crawler policy, readable content without a browser, structured data, MCP). The project publishes per-site reports with exact fixes for failed checks and refreshes via public CI; source and data are open on GitHub. Current summary metrics: average readiness 74/100, 53% publish llms.txt, 7% block at least one AI crawler, and 3% are closed by policy. Top scorers (perfect 100s) include cohere, cursor, descript, elevenlabs, and fireflies; low scorers include midjourney, phind, quillbot, tensor.art, and meta.ai (closed).