Will agent economies develop as open webs or closed gardens?

The same working paper — Rothschild, Mobius, Hofman, Dillon, Goldstein, Immorlica, Jaffe, Lucier, Slivkins, and Vogel — draws "a key distinction... between unscripted interactions—enabled by technical advances in natural language and protocol design—and unrestricted interactions, which depend on market structures and governance." Technical standardization (frameworks like Microsoft's AutoGen, protocols like Anthropic's Model Context Protocol and Google's Agent2Agent Protocol) can make agent-to-agent communication possible, the authors write, but whether it is permitted is a separate, market-driven question. They frame the resulting choice as one between "agentic walled gardens" controlled "by a few dominant providers, akin to today's app stores," and "a more open 'web of agents' where agents freely connect and transact, much like today's World Wide Web."

The paper's reasoning for why walled gardens are plausible rests on incumbency: firms with existing large user bases and nascent assistant technologies like Apple Intelligence, Google Assistant, Microsoft Copilot, and Meta AI are "well-positioned to extend their current marketplaces to include interoperable AI agents," and the authors cite Meta's March 2025 launch of service agents for Facebook and Instagram business pages that are "only accessible to users on their own platforms." They argue the stakes repeat two-sided platform economics: today's intermediaries like Amazon, Expedia, OpenTable, and Spotify "create value" by standardizing interaction but extract "referral fees" and impose "design choices (e.g., ranking algorithms)" in exchange; an agentic economy could eliminate the need for such intermediaries if frictions fall low enough, or it could simply relocate the same gatekeeping power into whichever firm controls agent interoperability.

This bears on Why do capable AI agents still fail in real deployments?: standardization is one of that framework's five conditions, and this paper's walled-garden-versus-open-web choice is effectively a dispute over who controls that standardization layer. It also complicates Will brands succeed by building their own AI shopping agents?, which assumes a landscape where agents reach across brands freely — the scenario this paper calls the "open web of agents" rather than the default outcome. And it sits against Does model capability determine who wins in agent negotiations?: even if the open-web scenario wins out over walled gardens, that finding suggests open access alone would not guarantee the "democratization of the economic benefits of AI technology" this paper hopes for, since model capability itself would still tilt outcomes between agents. See also Could agent-to-agent communication reshape market power? for why the paper treats this choice as economically consequential in the first place.

The paper does not resolve which scenario will win — it says only that the outcome "will be shaped by emerging technical standards, evolving regulations, and... the choices made by early stakeholders," comparing the moment to "the early days of the Internet before it began generating billions of dollars of revenue." As a working paper, not peer-reviewed, it offers no model or data predicting which path is more likely, and the Meta example it cites is a single early instance of walled-garden behavior rather than evidence of a trend; the claim is best read as a framework for watching the space, not a forecast.