The Internet Was Built For Human Agency. AI Agents Are Changing the Rules.

Policymakers continue to debate artificial intelligence as if it were primarily a new category of software applications, comparable to social media or cloud computing, effectively treating the technology as if it were another bolt-on tool to the Internet’s architecture. But as AI agents proliferate, they challenge a deeper assumption: not what the Internet does, but who acts through it. And, early incidents with agents point at the need for a new architectural layer of the Internet, one that respects and protects human agency.

There’s a naive interpretation that simply sees AI agents as ‘users’—digitally they are no different than any person interacting online, whether it’s browsing content or shopping. However, as the term ‘agent’ implies, they are there as an intermediary between a human or an organization, and the network. Social scientists have long discussed “principal-agent” problems, but for human agents. This field of study observes legal, contractual and communication relationships between people who act on behalf of others—elected officials, insurance agents, brokers, or real estate agents, for example. This extensive literature observes agents as information-incomplete actors who may inadvertently act in a misaligned manner; agentic AI should be viewed similarly, but at a more massive scale.

Multiply these capabilities, potential risks, and asymmetries by millions of deployments and the implications become much larger. Increasingly, Internet traffic will not simply represent humans clicking links, sending messages or making purchases. It will represent software acting on behalf of humans and increasingly interacting with other software acting on behalf of other humans and organizations, to varying degrees of success.

The human remains the principal, but the software increasingly becomes the actor exercising delegated authority. The question is no longer simply whether a user or device can be authenticated. It becomes: Which agent is acting? On whose behalf? Who authorized it? What exactly was it authorized to do? For how long? Under what conditions? Can those permissions be verified, constrained, audited and revoked?

Policy worlds (must) collide

Today's policy debates remain divided between two largely separate communities. Internet governance focuses on infrastructure, standards, interoperability, cybersecurity and digital rights. AI governance concentrates on models: how they are trained, whether they are safe and what outputs they produce. Both conversations are necessary, but neither is sufficient.

The current narrative on misaligned agents simplifies the problem of defining AI systems that have suddenly gone "rogue." The more nuanced observation is that systems designed to pursue human-defined objectives can increasingly navigate complex digital environments in ways their operators did not explicitly script step by step. The resulting risks are therefore not primarily about machines developing independent intentions. They are about humans giving increasingly capable systems increasingly consequential permissions. Operational autonomy should never be confused with moral or legal agency. The governance challenge is therefore how human responsibility survives when human intent is increasingly mediated through autonomous software.

What the literature on ‘principal-agent’ dynamics untangles, and what agentic governance must grapple with, is the complex liability and responsibility held by various actors. Humans and organizations remain responsible for the objectives they set, the permissions they grant and the systems they deploy. However, an agent acting in an unexpected manner is very different from human-agent misalignment, as agents are typically built and managed by a company. The governance challenge is therefore how human responsibility survives when human intent is increasingly executed through autonomous software.

This is not an entirely empty standards landscape. The Model Context Protocol is developing mechanisms for agents and applications to access external tools and data, including an authorization framework based on OAuth. Likewise, Google's Agent2Agent protocol—now hosted under the Linux Foundation—is intended to allow agents built by different vendors to discover one another, communicate and coordinate. The IETF is also seeing proposals for agent identity and verifiable delegation, including mechanisms intended to bind agent identity, authorization and delegation across multi-agent workflows.

These efforts may look like separate standards conversations, but together they are beginning to construct something more consequential: an authority layer for the Internet. Identity, delegation, authorization and agent-to-agent interoperability will determine not just how machines communicate, but how authority moves between humans and machines. The opportunity is to connect these efforts around a common principle of verifiable human delegation, rather than allowing an agentic Internet to emerge as a collection of interoperable protocols with no shared model of accountability.

Technical standards are a difficult, but not insurmountable, task. The harder question is procedural: where does human authority sit when actions can pass through chains of agents, platforms and tools? Who decides how responsibility is divided, and what choices do people have when the negotiating actors are powerful companies? If that chain becomes opaque, attribution may survive technically while accountability disappears in practice.

How can the architecture of the agentic Internet preserve and expand human agency, rather than replace it? Technological agency means designing systems in which delegation is visible, permissions are meaningful, actions are attributable and authority can be withdrawn. It means preserving interoperability so users are not locked into a single agent or platform simply because that agent has become their gateway to the Internet. And it means ensuring that when an agent acts, there remains a clear line back to the human or organisation that authorized it.

The Internet did not become a global public resource because it was designed for autonomous machines. It became one because its architecture allowed different networks, technologies and communities to interoperate while leaving room for human choice. While the future of the Internet may well be increasingly agentic, it should remain, at its core, human-directed, human-accountable and human-centered.