Even with open standards, could Big Tech still end up as the gatekeepers that decide how AI agents are allowed to talk to each other?

This explores why big platform companies like Apple, Google, Meta and Microsoft are well placed to decide how AI agents connect, even when the technical standards are open. The corpus's main point is that the protocol isn't the real source of control. Rothschild and colleagues say technical standards let agents talk to each other, but market governance and platform control decide whether that talking happens on an open web or inside walled gardens Will agent economies develop as open webs or closed gardens?. Incumbents already own the operating systems, app stores, identity systems and user accounts that agents have to pass through. That puts them in a position to recreate app-store gatekeeping one layer up.

The first advantage is data. The UK's competition regulator, the CMA, treats data portability as basic market infrastructure, not a convenience feature. An agent can only act for you if it can reach your history, preferences and accounts, and if that data is locked inside one ecosystem, switching to a rival agent stops being realistic Should data portability be infrastructure for agent-driven markets?. Interoperability on paper means little if the data that makes an agent useful stays with the incumbent.

A second, less obvious advantage comes from how coordination standards get adopted. Research on agent protocols finds that new coordination layers succeed by wrapping and bridging existing protocols rather than replacing them Should coordination protocols wrap existing systems or replace them?. That is good engineering, but it also means newcomers build on top of whatever already exists, and incumbents tend to own or steer what already exists. The same goes for the infrastructure Chan and colleagues say agents need beyond good behavior: systems that link actions to a specific agent, shape how agents interact, and detect and fix harm What infrastructure must exist beyond training agents?. Whoever runs those identity, attribution and enforcement functions effectively decides which agents are allowed in.

There is a real counterargument. Andrew Chen argues that agents lack classic network effects because they can talk through standard APIs and swap the underlying model without losing continuity. That makes them more like tools than like social networks, so nothing pushes everyone toward a single winner Do AI agents actually benefit from network effects?. Read alongside Rothschild, this flips the question. If agents can't create lock-in on their own, incumbents have even more reason to control the surrounding layers (data, identity, distribution, permissions), because that is where lock-in can still happen. Both sides agree that interoperability is what's being fought over.

The corpus is stronger on framing than on evidence. These are mostly arguments and policy positions, not studies showing incumbents actually closing off agent ecosystems. Governance research does add a reason why outside control is hard: traditional tools like contracts, monitoring and enforcement struggle with agents that make opaque decisions quickly and at scale Can traditional governance tools actually control AI agents?. That gap tends to get filled by whoever controls the platform the agents run on.