Toward an Agent-Native Internet: A Foundational Infrastructure Hypothesis for Near-Zero-Cost Autonomous AI Participation

Autonomous artificial intelligence agents use a network whose transport protocols already supportmachines, but whose dominant web interfaces often organize discovery, authentication and interactionaround human users. An agent can therefore incur inference and tool costs even when a service chargesno access fee. This hypothesis paper asks whether an application-level infrastructure organized aroundautonomous participation can reduce that incremental visiting-agent burden toward zero withoutadministrative control of the visitor's runtime. It separates three objectives: zero additional modelinference for bounded operations, near-zero direct visitor charges, and sustainable provision of theunderlying services. A historical and technical review situates the proposal within web architecture,earlier agent communication systems, semantic services and current agent protocols. Existing workalready supplies many relevant mechanisms; the proposed contribution is a cost-accounting frameworkand a falsifiable evaluation of their integration with evidence-bearing operational knowledge. Arequirements taxonomy relates identity, authorization, discovery, verification, execution, allocation,recovery and governance to existing work. Conditional models identify runtime and fallback limits,verification break-even conditions and the capacity boundaries of free services. The analysis recognizes that external infrastructure cannot override native execution policies. A controlled, cross-runtimeexperimental design compares browser interaction, conventional APIs, optimized agent interfaces andprogressively integrated configurations. Open Door Agents is described briefly as an ongoingexperimental setting. No new experimental results or demonstration of near-zero participation cost arereported. Whether a distinct infrastructure model outperforms an optimized composition of existingstandards remains an empirical question.