Orchestration
Three Agent Protocols, One Shared Blind Spot: Governance
A new arXiv paper examines MCP, A2A, and ACP — the three protocols most AI agent systems now rely on for tool access and inter-agent coordination — and asks what they structurally cannot express, not just what they haven't gotten around to yet.
The paper frames each protocol by its original design intent. MCP, introduced by Anthropic in late 2024, standardizes how agents access tools, data sources, and prompts through a client-server architecture, defining Tools, Resources, and Prompts as its core primitives. A2A and ACP were each built for different coordination concerns, and the researchers analyze five protocols total as representative of the major architectural approaches to agent interoperability as of mid-2026.
The useful move here isn't cataloguing features — it's treating governance as a first-class design question rather than something bolted on later. A protocol that handles task delegation well can still have no vocabulary for who's accountable when a delegated task produces a bad outcome, no mechanism for an agent to contest a decision made on its behalf, and no way to encode the kind of negotiated consent that human institutions rely on. Those aren't bugs you patch with a new field in a JSON schema; they're missing layers of the protocol's conceptual model.
This matters now because the industry has spent 2025 and 2026 converging on these three protocols as the de facto plumbing for multi-agent systems — A2A shipped v1.0 and joined the Agentic AI Foundation, MCP crossed into Linux Foundation stewardship with over 110 million monthly downloads. Convergence at the transport layer is happening fast. Convergence on how disputes, consent, and accountability get expressed within that transport layer is not happening at all, because none of the incumbent specs were built to carry that information in the first place.
The paper's framing draws on older governance theory — commons management, deliberative consensus — to ask what a protocol would need to actually support multi-party accountability rather than just multi-party message-passing. That's a harder problem than adding an audit log, because it requires agreeing on what counts as a legitimate claim, objection, or remedy between autonomous parties that don't share an institution.
The practical takeaway: teams building on MCP, A2A, or ACP today are getting fast, reliable message delivery between agents. They are not getting a shared language for resolving who was supposed to do what, who gets to object, or how a dispute gets adjudicated — and retrofitting that later, once millions of agents are already running on these rails, will be a much bigger lift than building it in now.