Agent-First Web: Redesigning the Web for AI Agents (arXiv paper)

Eranga Bandara and 20 coauthors submitted an arXiv paper (arXiv:2606.19116) on 17 Jun 2026 that argues the World Wide Web must be rebuilt for AI agents. The paper, a 5,733 KB submission on arXiv, lays out ten design principles and a three-layer redesign covering access, economic and content systems.

What does the paper propose?

The paper proposes a principled redesign across three layers: access, economic and content, and it bundles these into ten design principles. At the access layer the authors call for agents acting for humans to inherit equivalent access rights and to identify themselves using agent identification metadata in HTTP requests, analogous to browser headers; at the economic layer they propose an intent-based tier framework and a token-based subscription model; at the content layer they propose the Agent Text Markup Language (ATML), a four-level human supervision tier model, and a cryptographic provenance chain.

How would the access, economic and content layers change?

Access would add agent identification metadata in HTTP requests and a dual-layer architecture serving both human-readable and agent-optimized content from the same domain. The economic model would shift from pageview metrics to an intent-based tier framework that treats an agent's economic obligation as mirroring the human it represents, and would meter content with a token-based subscription system and a commissioned content economy. Content controls would address what the paper terms "epistemic recursion," the self-referential loop in which AI-generated content is consumed by agents to create more content, by introducing ATML with four supervision levels and a cryptographic provenance chain to trace origins.

How does the paper define the core problems it aims to solve?

The authors say the web was built on an assumption held for three decades: that the primary consumer of web content is a human being. That assumption, they argue, infects access models, economics that rely on human attention, and content designed for human perception. The rise of AI agents as intermediaries invalidates that assumption, and the paper documents how current responses—blanket blocking, CAPTCHA-based exclusion, and treating agent access as extraction—are inadequate and adversarial to legitimate agent use.

Why it matters

Putting agents first would renegotiate the web's social contract between publishers, users and intermediaries. The paper frames this as more than a technical change: it touches incentives and provenance, by tying agent behaviour to the human intent they represent and by proposing cryptographic provenance to preserve human-grounded truth. If adopted, the proposals would reshape publisher metering, how content is labeled and served, and the incentives for producing AI-generated versus human-commissioned work.

What to watch

Look for any early specifications or prototypes of ATML, and for experiments implementing agent identification metadata in HTTP requests and dual-layer hosting for human and agent content. Track whether the intent-based tier ideas or token-based subscription models appear in technical discussions, standards drafts or developer implementations following the paper's publication on arXiv.

The paper is available on arXiv as arXiv:2606.19116 (submitted 17 Jun 2026) and lists its full draft and supplementary files in the submission record.