Recoleta
GitHub
Source note

Governance Gaps in Agent Interoperability Protocols: What MCP, A2A, and ACP Cannot Express

Authors
Richard Kang, Yudho Diponegoro
Source
arXiv
Published
Collection
Software Intelligence
Agent InteroperabilityMulti Agent GovernanceProtocol AnalysisAgent CoordinationEnterprise AI
Open arXivMarkdown source

Summary

This paper argues that MCP, A2A, ACP, ANP, and ERC-8004 support agent coordination but lack protocol-native governance for agent communities. Its main claim is that governance needs a separate architectural layer above current interoperability standards.

Problem

  • Enterprises are deploying heterogeneous agent fleets that can discover tools, exchange messages, delegate tasks, and record trust signals, but these protocols do not define how agents make governed group decisions.
  • The missing capabilities matter for regulated or high-risk settings, such as production code changes, compliance review, and research arbitration, where decisions need membership control, deliberation, voting, dissent records, human escalation, and audit.
  • Without shared governance primitives, each application must implement its own decision process, which weakens interoperability and audit consistency.

Approach

  • The paper defines a six-part governance taxonomy: G1 membership, G2 deliberation, G3 voting, G4 dissent preservation, G5 human escalation, and G6 audit/replay.
  • It evaluates five protocols: MCP v1.1, A2A v1.0.1, ACP, ANP, and ERC-8004.
  • Each protocol-dimension pair is classified as Supported, Partial, or Absent based on what the protocol specification encodes, rather than what an application could build on top.
  • The authors separate gaps that could be addressed through protocol extensions from gaps that may need a new layer.

Results

  • Across the five protocols and six governance dimensions, no protocol fully supports any governance dimension in the reported matrix.
  • Voting, dissent preservation, and human escalation are Absent in all 5/5 protocols.
  • Coverage scores are low: MCP v1.1 scores 1/12, A2A v1.0.1 scores 1/12, ACP scores 2/12, ANP scores 0/12, and ERC-8004 scores 2/12.
  • Membership is Partial in A2A, ACP, and ERC-8004, but Absent in MCP and ANP.
  • Deliberation is Partial only in ACP and Absent in the other 4 protocols.
  • Audit/replay is Partial in MCP and ERC-8004, but Absent in A2A, ACP, and ANP; the paper says these partial cases come from session state or blockchain history rather than governance-specific replay semantics.

Link