Agent fitness becomes measurable at the environment boundary
Source → Observed → Interpretation → Model implication
Introducing the Agent Readiness score. Is your site agent-ready?
View source →Cloudflare Radar scanned 200,000 of the most visited domains after excluding categories where agent readiness was not relevant, and published adoption data for agent-facing standards by domain category.
Cloudflare reports that 78% of scanned sites expose robots.txt, 4% declare AI usage preferences using Content Signals, 3.9% support Markdown content negotiation, and fewer than 15 sites in the dataset expose either MCP Server Cards or API Catalogs.
The Agent Readiness scanner scores sites across four dimensions: Discoverability, Content, Bot Access Control, and Capabilities. Checks include robots.txt, sitemaps, Link headers, Markdown for Agents, Content Signals, Web Bot Auth, Agent Skills, API Catalogs, OAuth discovery, MCP Server Cards, and WebMCP.
For each failing readiness check, Cloudflare provides a prompt that can be given to a coding agent to implement support. The scanner is also available programmatically through an MCP server and URL Scanner API.
Cloudflare reports that serving Markdown rather than HTML reduced token usage by up to 80% in some cases, making agent consumption faster and cheaper.
In a documentation benchmark using Kimi-k2.5 through OpenCode, Cloudflare reports that the agent pointed at Cloudflare documentation consumed 31% fewer tokens and arrived at the correct answer 66% faster than the average comparison documentation site that had not been similarly refined for agents.
Cloudflare attributes poorer performance on other documentation sets to large context files that force repeated grep/search loops, fragmented context, additional thinking tokens, more requests, and higher latency and cost.
The record operationalizes environmental fitness rather than treating it only as an explanatory concept. Discoverability and agent-readable content affect the cost of acquiring sufficient context; callable capabilities and authentication metadata affect executable interaction; readiness scores and Internet-wide adoption data make environmental traits observable; token and latency results expose economic consequences; and failing checks can be turned directly into coding-agent remediation work. The documentation benchmark provides first-party evidence that changing the environment can improve agent efficiency without changing the agent itself.
REFINES. Selection can be made measurable and intervention-driven. Environmental traits can be defined, assessed across a population, exposed as gaps, changed, and reassessed. This suggests a concrete Selection loop: define desirable environmental traits, measure their presence, intervene where the environment is weak, and then evaluate whether agent cost or task performance improves. The evidence also sharpens the model's claim that the environment belongs in the fitness function by attaching measurable context, latency, and token costs to different environmental designs.
What this does not establish
- The 200,000-domain scan measures adoption of selected standards and readiness checks; it does not measure end-to-end agent task success across those domains.
- The 31% token reduction and 66% faster correct-answer result are Cloudflare's first-party benchmark and are not an independently replicated causal study of every readiness check in the score.
- The source does not establish that the Agent Readiness score predicts agent engineering success across arbitrary agents, tasks, repositories, or domains.
- Some scored standards are emerging, and the source does not establish which of them will achieve broad adoption or remain durable parts of the agentic Web.
- The source does not establish Cooperation, Specialization, 1,000-agent operation, or the model's proposed evolutionary sequence.
Which measurable environmental traits best predict an agent's cost and probability of reaching sufficient context and completing useful work?