The Internet becomes part of the agent fitness function
Source → Observed → Interpretation → Model implication
Building an open Agentic Internet: readable, discoverable, callable, and payable
View source →Cloudflare reports that well-behaved bots generate billions of requests re-fetching pages that have not changed, creating machine effort and cost without a useful outcome for either side.
Cloudflare describes agents as a new class of Internet actor and argues that the human-oriented web creates unnecessary context, interaction, and economic friction for them.
Markdown for Agents is presented as a server-side way to reduce bandwidth, token use, and irrelevant HTML in agent context windows; AI Search is presented as an agent-oriented discovery interface and AEO as a way for resource owners to measure whether agents can discover them.
WebMCP lets sites expose explicit tool contracts to agents instead of requiring DOM parsing and synthesized clicks, while also providing clearer signals to content owners about what resources agents actually use.
Web Bot Auth and PACT are presented as identity and access primitives that let agents identify themselves and let sites decide which agents they trust or admit.
Cloudflare presents x402-compatible payments, wallets, and receipts as a way for agents to transact directly with sites and for both sides to account for what was used and paid for.
Cloudflare says these mechanisms are intended to help agents and domain owners cooperate on the open Internet rather than collide, and emphasizes open standards including MCP, x402, Web Bot Auth, and PACT.
The source extends the Selection hypothesis beyond repositories and engineering organizations. When agents interact with infrastructure designed for humans, the environment imposes measurable context, execution, and economic waste. Cloudflare's proposed agent-native interfaces make those environmental constraints more explicit through machine-readable content, discovery, identity, callable contracts, observability, and transaction rails. The source also points toward Cooperation between independent agents and domain owners, but does not yet demonstrate that this cooperation provides a durable fitness advantage over ad-hoc interaction.
REFINES. Selection pressure may be scale-recursive. As agent populations grow, the relevant fitness environment can move outward from the agent configuration to the repository, organization, platform, and shared Internet protocols. Agent-native standards may create preconditions for Cooperation by reducing ambiguity and coordination cost across organizational boundaries, but the source does not yet establish a durable cooperation advantage.
What this does not establish
- It does not establish that Cloudflare's proposed agent-native protocols outperform conventional web interaction across repeated real-world agent tasks.
- It does not establish a durable Cooperation advantage relative to isolated or ad-hoc interaction.
- It does not establish stable Specialization or a higher-order agentic engineering object.
- Cloudflare's architectural and product claims are first-party. The reported redundant-request traffic establishes environmental friction, not the effectiveness of every proposed remedy.
- The source concerns Internet-scale agent interaction broadly; coding agents are identified as a mature form of this actor, but the reported billions of requests are not specifically attributed to coding agents.
As agent populations grow across organizational boundaries, which shared protocols reduce context, coordination, and economic cost enough for cooperation between independent agents and systems to become fitter than ad-hoc interaction?