The Internet of Agents Will Be Built by Platforms That Own the Protocol

A number caught my attention when I read the Gartner report on multi-agent systems this week. Gartner describes three phases of MAS evolution: single-platform, cross-platform, and the Internet of Agents. By 2028, they predict 60% of multi-agent systems will support multivendor interoperability. The first two phases are about scaling agents inside a single environment. Phase 3 is different. The Internet of Agents is a global network where agents discover each other, negotiate tasks, and coordinate work using shared protocols. That sounds like a technical milestone. I think it is a platform governance event, and the IS literature has been describing exactly how it will play out for years.

The phase transition from cross-platform to Internet of Agents requires a protocol. Agents built by different vendors need a common language to discover each other, negotiate task allocation, verify credentials, exchange data, and settle transactions. The company that defines that protocol controls the ecosystem. This is not a technical insight. It is a structural insight about how platforms build power. A platform is an architecture with a core and a periphery. The owner controls the core interfaces, the APIs, the terms of participation, and the rules that govern coordination. Complementors build in the periphery, and the more they build, the more valuable the platform becomes. The protocol is the ultimate boundary resource. It determines which agents can discover each other, what information they can exchange, how trust is established, and what fees the platform can extract from every interaction.

I wrote about this earlier in my post on platform governance and multi-sided markets. The mechanism that matters here is platform envelopment. Platform envelopment happens when a platform in one market uses its governance position to enter an adjacent market, bringing its user base and ecosystem along. Search engines became portals. Social networks became advertising exchanges. The App Store became a marketplace for services, not just software. The Internet of Agents will be the next envelopment target. The platform that controls the agent protocol will extend into every industry where autonomous coordination between digital agents creates value. Logistics, healthcare, finance, supply chain, enterprise workflows. Each one becomes a new market accessed through the same governance layer.

Mayer, Kostis, Strich, and Holmstrom (2025) studied how GenAI as a boundary resource reshapes platform governance. They found that it creates three specific challenges: validation problems because outputs need checking, standardization problems because generative outputs resist stable platform rules, and complementor-skill problems because participants need new capabilities to work with unpredictable AI. The Internet of Agents compounds every one of these challenges. When the complementor is not a human developer but another AI agent running on a different platform, validation becomes harder because the orchestrator cannot inspect the agent's internal logic. Standardization becomes harder because agents may interpret the same protocol in different ways. The skill problem shifts from human training to agent alignment. The platform that owns the protocol can set the terms for all three.

Apple is the precedent I keep returning to. Before the iPhone, mobile software was distributed through carrier deals and physical media with almost no ongoing governance. Apple built the App Store as the single distribution channel and wrote rules that governed every transaction. Developers accepted the terms because the user network was too valuable to leave. When the European Union passed the Digital Markets Act and forced Apple to open the App Store to alternative payment systems and sideloading, Apple responded with fee structures and warning screens designed to make alternatives feel risky and inconvenient. The walled garden did not collapse. It got a door that most users will never walk through. The people I talk to who build on iOS know this. They stay because the network effects are asymmetric. Users do not leave for a more open platform because users do not care about openness. They care about the apps they already have and the people they already communicate with.

Microsoft is making the same bet with the Copilot ecosystem, and I think it is a smarter long-term play than most analysts give it credit for. Copilot is not a chatbot feature. It is an infrastructure play for agent orchestration inside the enterprise. Microsoft is building the connectors, the authentication layer, the data classification framework, the monitoring tools, and the compliance graph that agents will need to work across enterprise systems. When an agent needs to pull a record from Dynamics, draft a response in Word, check it against retention rules in Purview, and send it through Outlook, every step touches a Microsoft boundary resource. The agent never leaves the Microsoft ecosystem because the ecosystem already covers the full workflow. Baird and Maruping (2021) showed that agentic IS changes the relationship from use to delegation, with three mechanisms: appraisal, distribution, and coordination. Microsoft is embedding its protocol into all three. Appraisal happens through Azure AI content safety. Distribution happens through Copilot's task routing. Coordination happens through Graph connectors and Teams. The switching costs are built into every layer.

I am not saying Microsoft will win because its technology is superior. I am saying Microsoft will win because its platform position is already superior. The installed base of Microsoft 365 users, the enterprise contracts that renew annually, the compliance certifications that take years to replicate, and the developer habit of building on Azure create a multi-sided network that no agent protocol startup can match in this decade.

Open protocols demonstrate the difficulty of competing against this logic. ActivityPub is a decentralized protocol for social networking. It works. Mastodon runs on it. Threads adopted it. But Mastodon has never approached the scale of the proprietary platforms. The reason is not technical quality. The platform owner extracts value by controlling the coordination layer. An open protocol distributes that value to everyone, which means no single party has the incentive to invest in making it better, marketing it, or building the support infrastructure that enterprise adoption requires. Open protocols survive as alternatives that technically sophisticated or ideologically motivated users choose. They do not unseat proprietary platforms because most users optimize for convenience and network size, not for protocol openness.

The next three years will test this pattern in a new domain. The Internet of Agents will not be an open internet where any agent can negotiate with any other agent on equal terms. I think it will be three or four walled gardens, connected to each other through grudging interoperability that the platform owners implement only because customers or regulators demand it. Google will have its agent protocol and it will work best with Google services. Microsoft will have its protocol and it will work best inside the Microsoft 365 graph. OpenAI will have its protocol and it will work best with ChatGPT and the agents that build on its models. A fourth might come from Apple or from a consortium of financial institutions that cannot afford to depend on a single Big Tech protocol.

The interoperability that Gartner predicts will happen. Sixty percent of multi-agent systems will support multivendor communication by 2028. But interoperability is not openness. APIs are not level playing fields. The platform literature has been telling us this since the first research on boundary resources and governance. The owner of the protocol sets the terms. The complementors adapt. The walled garden persists even when the walls have doors, because the doors are designed to be expensive to walk through and the garden was designed to be hard to leave.