The web is hostile to bots, and a new open experiment called Agentability is proving it with receipts. Launched as a daily show on Hacker News, the project deploys an AI agent to complete ten real-world web errands every morning, publishing every transcript verbatim. The goal is radical transparency: no retries, no editing, and no cherry-picking. The agent, powered by deepseek-flash, operates under strict constraints—read-only HTTP GET requests with no JavaScript execution, no logins, and no human intervention. This setup mirrors the worst-case scenario for current web infrastructure, where dynamic content and authentication walls frequently block automated readers.

The Daily Grind: 8/10 Success Rate

On October 6, 2026, the agent managed to complete 8 out of 10 errands, a surprisingly high success rate given the constraints. However, the underlying metrics reveal the friction involved. The agent hit 10 bot walls, read 137 pages, and visited 43 distinct sites to gather the necessary information. The errands are generated by an AI producer using deepseek-v4-pro, which scans live search trends to formulate questions like sports schedules, earthquake magnitudes, and product costs. The project’s philosophy is that agent capability is an empirical question, best answered by exposing the raw logs of success and failure to the public.

Scoring the Web: 113 Sites Audited

Beyond the daily errands, Agentability maintains a standing index of 113 well-known websites, audited weekly for agent readiness. The current average readiness score across these sites is 74/100. The audit checks for specific technical conventions that help agents navigate the web, including the presence of llms.txt files, crawler policy compliance, structured data, and MCP (Model Context Protocol) support. The data shows that only 53% of these major sites publish an llms.txt file, while 7% actively block at least one AI crawler, and 3% are closed to AI by policy. This index serves as a diagnostic tool, linking agent failures in the daily show to specific technical deficiencies on the target sites.

The Good, The Bad, and The Blocked

The leaderboard highlights a stark divide in web accessibility. Cohere, Cursor, Descript, ElevenLabs, and Fireflies.ai all achieved a perfect 100/100 grade, largely due to their adoption of llms.txt files. These sites are effectively agent-friendly. In contrast, the bottom of the table features sites like Midjourney, Phind, QuillBot, and Tensor.art, all scoring a dismal 15/100. Meta.ai sits at the very bottom with a 10/100, marked explicitly as 'Closed by policy.' The project argues that when an agent fails during a daily errand, the index has usually already predicted the obstacle, creating a feedback loop between real-world performance and technical auditing.

Key Takeaways

  • Agentability publishes verbatim, unedited transcripts of AI agents attempting web tasks, emphasizing transparency over curated success.
  • The project uses a strict 'no JavaScript, no login' constraint to test the raw accessibility of the web for simple HTTP agents.
  • A parallel audit of 113 major sites reveals that only 53% currently support llms.txt, a key convention for agent navigation.
  • Top performers like Cohere and Cursor score 100/100, while Meta.ai scores 10/100 due to explicit policy blocks.
  • The project provides actionable fixes for site owners, linking failed checks in the index to concrete technical recommendations.

The Bottom Line

Agentability is doing the dirty work of exposing the web's infrastructure gaps. If your site isn't in the top 100, you're invisible to the next generation of users. For developers and product owners, this is a wake-up call: the web is becoming a machine-readable interface, and those who ignore llms.txt and structured data are building digital walls around their content. The 'agentability' score is no longer just a vanity metric; it’s a measure of reach in an increasingly agentic world.