Agent Readiness Compare

File checks and task runs: two ways to test agent readiness

One tells you the signals are present. The other tells you whether an agent got through.

In short

A file check confirms that something exists: robots.txt returns 200, llms.txt has an H1, a page answers a markdown request. A task run sends an agent to find a price or reach a demo request and records where it stopped. You need both. File checks are fast and repeatable. Task runs show whether the signals add up.

What a file check tests

A file check makes one request and inspects the response. Typical examples: robots.txt returns HTTP 200 and parses, /llms.txt exists with an H1 and working links, a page returns markdown when asked, /.well-known/agents.json returns JSON, and an unpaid request to a paid endpoint returns HTTP 402. Each has a clear pass or fail, which makes file checks cheap enough to run on every release. Scanners such as Ora and Cloudflare's are mostly file checks at scale.

What a task run tests

A task run gives an agent an intent, such as finding the price of a team plan, and follows it through the site. Ora's Journey does this: it points a real agent at a domain with an intent such as "find pricing" and records every step, stall and outcome. Curated tasks run without a key. Free-text tasks need a partner key, and each domain is capped at 100 runs per rolling 24 hours. is-agentic.com attaches one observed agent journey to each report as evidence, outside the numeric score.

Task runs catch failures no single file check sees: a price that is in the HTML but hidden behind a toggle, a sign-up form whose fields have no labels, or a bot challenge on the last step.

Why the two sometimes disagree

A task depends on several layers at once. A site can pass every discovery and reading check and still fail the task because the demo form needs a custom gesture. The reverse happens too. An agent can sometimes finish a task on a site without an llms.txt, because the pages themselves are readable.

Ora's lab found (19 Oct 2026) that llms.txt, JSON-LD, sitemaps and markdown mirrors did not change retrieval in a 190-probe experiment on one docs site. Only a JavaScript-only answer and a 403 for agent user agents did. That is a vendor claim from a single site, but it shows why a present file is not a finished task. Treat a disagreement as information: it points at the step where the signals and the journey part ways.

Tasks worth running

  • Can an agent find and accurately explain our pricing?
  • Can it recommend the right plan for a stated requirement?
  • Can it find the documentation page that answers a buyer's question?
  • Can it complete a demo request or the next step we care about?

Keep the list short and the same from run to run, so results can be compared.

How often to run each

File checks belong in CI or a scheduled job. Ora's npx ax audit <domain> scores a URL from the terminal and prints fixes ranked by impact, so it can run in a pipeline. Task runs take longer and vary more between runs. Run them after releases that touch pricing, sign-in, bot rules or checkout, and keep a dated log.

Sources

  1. Ora, home page. https://ora.ai/, read 2026-11-04.
  2. Ora, llms-full.txt. https://ora.ai/llms-full.txt, read 2026-11-04.
  3. is-agentic.com. https://is-agentic.com/, read 2026-11-04.
  4. Ora, AgentReady launch post. https://ora.ai/blog/agentready-launch, read 2026-11-04.
  5. RFC 9309, Robots Exclusion Protocol. https://www.rfc-editor.org/rfc/rfc9309.html, read 2026-09.
  6. llmstxt.org. https://llmstxt.org/, read 2026-09.
  7. agents.json repository. https://github.com/wild-card-ai/agents-json, read 2026-09.
  8. x402.org. https://www.x402.org/, read 2026-09.

Last reviewed 2026-11-05