An academic taxonomy of MCP (Model Context Protocol) and A2A (Agent to Agent), ACP (Anthropic Computer Protocol), ANP (Agent Network Protocol), and ERC 8004 (Ethereum Request for Comments 8004), the five leading protocols for AI agent communication, finds that voting and dissent…
The protocols stitching together the enterprise AI agent boom were built to coordinate tasks, not to govern communities of agents making collective decisions. A new academic taxonomy of what five leading interoperability frameworks cannot express finds a systematic gap. Voting, dissent preservation, and deliberation are absent across every protocol surveyed, while the enterprises deploying heterogeneous agent fleets are running ahead of the governance layer that would make those fleets accountable.
The preprint "Governance Gaps in Agent Interoperability Protocols", submitted to arXiv on 30 June 2026 by researchers cataloguing multi-agent systems, applies a six-dimension governance framework to MCP, A2A, ACP, ANP, and ERC-8004. The dimensions, membership, deliberation, voting, dissent preservation, human escalation, and audit/replay, derive from organizational theory and enterprise governance standards rather than from networking protocols. The result is a structural map of where today's agent-to-agent plumbing ends and where collective-decision plumbing would need to begin.
The gap matrix is stark. Voting is uniformly unsupported: none of the five protocols encode a way for agents in a group to register a position, tally a result, or record how a decision was reached. Dissent preservation fares no better. There is no primitive for capturing that an agent disagreed with a collective outcome, or why. Deliberation, the structured exchange that produces a decision before it is made, is absent in most of the protocols and at best partial where it appears. Only human escalation, the ability for a person to step in, trends toward supported across the field. Membership and audit and replay sit in between, mostly partial.
The mismatch is not an oversight so much as an inheritance. The protocols were designed for task-oriented coordination: identity, capability discovery, tool access, and message exchange. The Model Context Protocol (MCP), introduced by Anthropic in late 2024 for connecting models to data sources and tools, carries that framing in its current specification: connections, requests, resources. Google's Agent-to-Agent protocol (A2A), now backed by more than 150 organizations according to the Linux Foundation, targets similar ground. Agent cards describe what an agent can do, tasks are delegated, message parts are exchanged. None of these primitives presupposes that a group of agents might need to disagree, vote, or produce a record that survives a future audit.
The deployment context, however, has moved. The motivating scenario the paper describes is enterprises running heterogeneous fleets of agents, for procurement, customer service, code review, and financial workflows, that increasingly need to make collective decisions under governance constraints that humans will be asked to defend. It is not a single agent calling a single tool. It is a community of agents whose decisions will need to be inspected, explained, and, where they go wrong, undone.
That is where the taxonomy earns its keep. By mapping each protocol against six dimensions, the analysis gives enterprise architects and auditors a shared vocabulary for naming what is missing. A team adopting MCP or A2A today can ask, with the paper in hand: how does this protocol preserve dissent? How does it support human escalation? What does audit and replay look like? The honest answer, for every protocol surveyed, is "not yet" or "not in the spec." The next move belongs to the standards bodies, the open-source maintainers, or the consortia that have staked their roadmaps on agent interoperability becoming an enterprise-grade layer.
A few caveats sharpen the picture. The paper is an arXiv preprint, not peer-reviewed, so the six-dimension taxonomy and the gap matrix are author claims that should be tested against independent enterprise governance standards before they are treated as settled. The adoption numbers behind A2A, 150-plus organizations and major cloud platforms, come from the Linux Foundation's own press release, which has an obvious interest in the protocol's momentum; corroborating those figures from buyer-side or analyst sources would strengthen any load-bearing claim. And the paper covers five protocols, not three as its title suggests. ANP and ERC-8004 sit alongside MCP, A2A, and ACP in the actual analysis.
What to watch next is specific. The pressure point is whether any of the five protocols adds a governance extension, such as a dissent channel, a vote primitive, or an audit log, before enterprises begin treating multi-agent decisions as production-grade records that regulators, courts, or auditors may need to inspect. The taxonomy is now on the table. The question is who builds the first governance layer on top of it.