Agentability Tests Whether AI Agents Can Use the Real Web

Summary

Agentability is an open experiment that tests whether AI agents can complete ordinary information-seeking errands on the public web. Each day, an AI producer turns current searches into ten tasks, such as finding a sports broadcast, measuring an earthquake, checking a price, or determining what happened. A DeepSeek-powered agent attempts the tasks using only plain HTTP requests, with no login, JavaScript, forms, retries, or human assistance; live search is included. The project publishes every transcript verbatim and reports both successes and failures. On the October 6, 2026 episode, the agent completed eight of ten errands, read 137 pages, visited 43 sites, and encountered ten bot walls, while two tasks were abandoned. Agentability also audits 113 well-known sites weekly against agent-oriented signals including llms.txt, crawler policy, readable content, structured data, and MCP. The index reports an average readiness score of 74 out of 100; 53% of sites publish llms.txt, 7% block at least one AI crawler, and 3% are closed to AI by policy. The project says its reports, open data, and reproducible checks are intended to show where the web works for agents and where sites create access barriers.