Paper deep dive
Governance Gaps in Agent Interoperability Protocols: What MCP, A2A, and ACP Cannot Express
Richard Kang, Yudho Diponegoro
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 98%
Last extracted: 7/5/2026, 1:33:42 AM
Summary
The paper presents a systematic gap analysis of current agent interoperability protocols (MCP, A2A, ACP, ANP, and ERC-8004) against a six-dimension governance requirements taxonomy. The authors argue that while existing protocols effectively handle task-oriented coordination (tool access, discovery, routing, and trust), they lack the primitives necessary for governed agent communities. The analysis reveals that dimensions such as voting, dissent preservation, and human escalation are universally absent across all analyzed protocols. The paper concludes that agent community governance constitutes a missing architectural layer (Layer 4) above current interoperability standards.
Entities (12)
Relation Signals (4)
Model Context Protocol (MCP) â lackscapability â Voting
confidence 100% ¡ G3 Voting: Absent. No voting primitives exist.
Agent-to-Agent Protocol (A2A) â lackscapability â Dissent Preservation
confidence 100% ¡ G4 Dissent: Absent. No dissent semantics in task responses or any extension.
ERC-8004 â partiallysupports â Audit/Replay
confidence 100% ¡ G6 Audit: Partial. On-chain transactions are inherently tamper-evident... However, the protocol does not define structured event types for decision reconstruction.
ACP â partiallysupports â Deliberation
confidence 100% ¡ G2 Deliberation: Partial. ACPâs negotiation patterns... constitute structured exchange with some challenge/response semantics.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Agent interoperability protocols (MCP, A2A, ACP, ANP, and ERC-8004) have rapidly matured to enable identity, capability discovery, tool access, and message exchange between autonomous agents. However, as enterprises deploy heterogeneous agent fleets that must make collective decisions under governance constraints, a question arises: can these protocols support governed agent communities, or only task-oriented coordination? We present a systematic gap analysis applying a six-dimension governance requirements taxonomy (membership, deliberation, voting, dissent preservation, human escalation, and audit/replay) derived from organizational theory, multi-agent systems literature, and enterprise governance standards. We analyze each protocol's specification against this taxonomy, classifying capabilities as Supported, Partial, or Absent. The resulting gap matrix reveals that voting and dissent preservation are universally absent across all five protocols, deliberation is absent or at most partial, and no protocol encodes the full set of primitives required for governed agent communities. We distinguish extensible gaps (addressable through protocol extension mechanisms) from structural gaps (requiring a new architectural layer) and assess time-sensitivity based on observed protocol evolution velocity. The analysis establishes that agent community governance constitutes a missing architectural layer above current interoperability standards, not a missing feature within them.
Tags
Links
- Source: https://arxiv.org/abs/2606.31498v1
- Canonical: https://arxiv.org/abs/2606.31498v1
Trouble viewing inline? Open PDF directly â
Full Text
32,182 characters extracted from source content.
Expand or collapse full text
Governance Gaps in Agent Interoperability Protocols: What MCP, A2A, and ACP Cannot Express Dr. Richard Kang DoiT International richard@doit.com Yudho Diponegoro AbstractâAgent interoperability protocolsâMCP, A2A, ACP, ANP, and ERC-8004âhave rapidly matured to enable identity, capability discovery, tool access, and message exchange between autonomous agents. However, as enterprises deploy heteroge- neous agent fleets that must make collective decisions under governance constraints, a question arises: can these protocols support governed agent communities, or only task-oriented coordination? We present a systematic gap analysis applying a six-dimension governance requirements taxonomyâmembership, deliberation, voting, dissent preservation, human escalation, and audit/replayâderived from organizational theory, multi-agent systems literature, and enterprise governance standards. We analyze each protocolâs specification against this taxonomy, classifying capabilities as Supported, Partial, or Absent. The resulting gap matrix reveals that voting and dissent preservation are universally absent across all five protocols, deliberation is absent or at most partial, and no protocol encodes the full set of primitives required for governed agent communities. We distinguish extensible gaps (addressable through protocol extension mechanisms) from structural gaps (requiring a new architectural layer) and assess time-sensitivity based on observed protocol evolution velocity. The analysis establishes that agent community governance constitutes a missing architectural layer above current interoperability standardsânot a missing feature within them. Index Termsâagent interoperability, governance, multi-agent systems, MCP, A2A, protocol analysis, agent communities, delib- eration I. INTRODUCTION The proliferation of LLM-based agents in enterprise en- vironments has driven rapid development of interoperability protocols. The Model Context Protocol (MCP) [1] enables agents to access tools and data sources. The Agent-to-Agent protocol (A2A) [2] standardizes discovery and delegation be- tween agents. The Agent Communication Protocol (ACP) [3] formalizes structured message exchange. The Agent Network Protocol (ANP) [4] provides graph-based routing with de- centralized identity. ERC-8004 [5] encodes on-chain identity, reputation, and validation registries. Together, these protocols address a coherent set of coordina- tion concerns: identity, capability declaration, discovery, tool access, message passing, and reputation. Enterprises deploying agent fleetsâAWS reports AgentCore customers scaling to 17 production agents within a year [6]âcan now reasonably expect their agents to find each other, exchange messages, and invoke each otherâs capabilities. Yet coordination is not governance. When a bank must decide whether autonomous coding agents should modify production systems, when a pharmaceutical company must arbitrate between competing research hypotheses, or when a regulatory body must determine whether an AI system meets compliance thresholds, the question is not âwhich agent can perform this task?â but âhow should agents collectively decide what to believe, test, or do?â This requires membership (who participates), deliberation (how claims are exchanged and challenged), voting (how positions are resolved), dissent preservation (how minority views survive), human escalation (when human authority is invoked), and audit (how the process is reconstructable). This paper asks: Do current agent interoperability protocols encode these governance capabilities? We present three contributions: 1) A governance requirements taxonomy comprising six dimensions (G1âG6) derived from organizational theory, multi-agent systems research, and enterprise governance standards, specifying what protocol-level primitives are necessary for governed agent communities. 2) A systematic gap matrix applying the taxonomy to five protocols (MCP v1.1, A2A v1.0.1, ACP, ANP, ERC- 8004), classifying each protocolâdimension pair as Sup- ported, Partial, or Absent based on specification-level evidence. 3) An extensibility and time-sensitivity assessment distin- guishing gaps that are addressable through existing exten- sion mechanisms from those requiring a new architectural layer, and characterizing the velocity at which the gap is narrowing. The remainder of the paper is organized as follows. Sec- tion I introduces the five protocols and their design intent. Section I derives the governance requirements taxonomy. Section IV presents the gap analysis and matrix. Section V discusses extensibility, time-sensitivity, and implications. Sec- tion VI positions against related work. Section VII concludes. arXiv:2606.31498v1 [cs.MA] 30 Jun 2026 I. BACKGROUND: AGENT INTEROPERABILITY PROTOCOLS We analyze five protocols representing the major archi- tectural approaches to agent interoperability as of mid-2026. Each was designed for a specific coordination concern; under- standing these design intents is necessary to interpret the gap analysis fairly. A. Model Context Protocol (MCP) MCP [1], introduced by Anthropic in late 2024, standard- izes how AI agents access tools, data sources, and prompts through a client-server architecture. The protocol defines three primitive typesâTools, Resources, and Promptsâexposed by MCP servers and consumed by MCP clients (typically LLM- based agents). MCP v1.1 (schema dated 2025-11-25) supports sessions, elicitation, sampling, and streaming. The protocol is tool-centric by design: it answers âwhat can an agent do?â rather than âhow should agents interact with each other?â MCP has achieved broad adoption, with over 1,000 community inte- grations and native support in AWS AgentCoreâs Gateway [6]. B. Agent-to-Agent Protocol (A2A) A2A [2], developed by Google and contributed to the Linux Foundation in 2026, enables agents to discover each other via Agent Cards (JSON-LD metadata describing capabilities, skills, and endpoints), delegate tasks, and exchange messages. Version 1.0.1 (May 2026) introduced an extension mechanism supporting ânew data, requirements, RPC methods, and state machinesâ [7]. A2A is delegation-centric: it answers âwhich agent can handle this task?â Four official example extensions exist (Secure Passport, Timestamp, Traceability, Agent Gate- way Protocol); none addresses governance. C. Agent Communication Protocol (ACP) ACP [3], developed by IBM Research, formalizes struc- tured multi-agent communication with negotiation semantics. ACP defines agent roles, message types, and negotiation patterns, drawing on FIPA-ACL heritage. It supports multi-turn dialogue with typed performatives (propose, accept, reject, counter). ACP is communication-centric: it answers âhow do agents exchange structured messages?â D. Agent Network Protocol (ANP) ANP [4] provides graph-based routing for agent networks using W3C Decentralized Identifiers (DIDs) for agent identity. ANP focuses on routing messages through agent networks without centralized registries. It is routing-centric: it answers âhow do messages reach the right agent across a network?â E. ERC-8004: Trustless Agents ERC-8004 [5], an Ethereum Improvement Proposal (Draft, created August 2025), defines three on-chain registries: an Identity Registry (agent addresses and metadata), a Repu- tation Registry (giveFeedback()/revokeFeedback() with signed fixed-point scores and tags), and a Validation Registry (independent validator attestations via TEE oracles, zkML verifiers, and stake-secured re-execution). ERC-8004 is trust-centric: it answers âwhich agents can be trusted?â It explicitly scopes itself to âdiscover, choose, and interact with agentsâ [5]. F. Design Intent Summary TABLE I PROTOCOL DESIGN INTENTS ProtocolFocusCore question answered MCP v1.1Tool accessWhat can an agent do? A2A v1.0.1DelegationWhich agent handles this task? ACPCommunication How do agents exchange messages? ANPRoutingHow do messages reach the right agent? ERC-8004TrustWhich agents can be trusted? None of these design intents targets the question: How should agents collectively govern community decisions? I. GOVERNANCE REQUIREMENTS TAXONOMY We derive a governance requirements taxonomy specifying what protocol-level primitives are necessary for governed agent communities. The taxonomy draws on three bodies of literature. Organizational theory. Habermasâs communicative ratio- nality [8] identifies structured argumentation, reciprocal chal- lenge, and consensus formation as prerequisites for legitimate collective decision-making. Parliamentary procedure [9] for- malizes membership (quorum), structured debate (motions, amendments), voting (majority rules, recorded dissent), and escalation (point of order). These map directly to protocol primitives. Multi-agent systems research. Ostromâs institutional anal- ysis framework [10] identifies governance rules for common- pool resources: boundary rules (membership), position rules (roles), choice rules (decision procedures), and information rules (transparency). Sierra et al.âs electronic institutions [11] formalize agent societies with norms, roles, and protocols. Recent work on LLM agent governance [12], [13] confirms these dimensions remain relevant for modern agent systems. Enterprise governance standards. Regulatory frameworks including SR 11-7 [14], ISO/IEC 42001 [15], and the EU AI Act [16] require auditability, human oversight, and ac- countability for AI system decisions. These translate to human escalation and audit requirements at the protocol level. From this synthesis, we derive six governance dimensions: A. Sufficiency Argument We argue these six dimensions are necessary and sufficient for governance (not for all coordination). Related concerns map to these dimensions or fall outside governance scope: ⢠Norm enforcement is a mechanism within G2 (deliber- ation) and G3 (voting)ânorms are enforced through the deliberative process. TABLE I GOVERNANCE REQUIREMENTS TAXONOMY (G1âG6) DimensionDefinition G1MembershipProtocol encodes admission, invitation, re- moval, and role assignment for commu- nity participants G2DeliberationProtocol encodes structured argument ex- change with turn-taking, challenge, and response semantics G3VotingProtocol encodes preference aggregation with quorum, rounds, and position reso- lution G4Dissent preser- vation Protocol ensures minority positions are retained in decision outputs, not silently dropped G5Human escala- tion Protocol defines conditions and mecha- nisms for routing decisions to human au- thority G6Audit/replayProtocol produces tamper-evident event logs enabling deterministic reconstruction of the decision process ⢠Reputation/trust is a prerequisite for governance (âwho is credible?â) but not itself a governance primitiveâit is already addressed by ERC-8004. ⢠Resource allocation is an outcome of governance deci- sions, not a governance mechanism. ⢠Incentive alignment/payment operates at a different architectural layer (economic coordination, not decision governance). B. Classification Criteria For the gap analysis in Section IV, we classify each protocolâdimension pair using three levels: ⢠Supported: the protocol specification explicitly defines primitives that satisfy the dimensionâs full definition. ⢠Partial: the protocol contains constructs addressing a subset of the dimensionâs requirements but not satisfying the full definition. ⢠Absent: the protocol specification contains no constructs addressing this dimension. Classification is based on what the specification encodes, not what could theoretically be built on top. This distinction is essential: any protocol can serve as transport for governance messages, but we assess whether governance semantics are protocol-native. IV. GAP ANALYSIS We now apply the taxonomy to each protocol. For each Partial classification, we specify which subset of the dimension is addressed and which remains absent. A. MCP v1.1 G1 Membership: Absent. MCP defines clients and servers but not community membership. There is no admission, invi- tation, or removal primitive. An MCP server either exists or does not; there is no concept of âjoiningâ or âbeing admitted toâ a group. G2 Deliberation: Absent. MCP enables tool invocation, not structured argument exchange. Sampling (server-initiated LLM calls) exists but carries no deliberation semantics. G3 Voting: Absent. No voting primitives exist. G4 Dissent: Absent. No dissent semantics exist. G5 Human escalation: Absent. MCPâs Elicitation feature (protocol version 2025-06-18) allows servers to request human input during tool execution, but this is user-input solicitation, not governance escalation. There is no protocol-level mech- anism for routing community decisions to human authority based on confidence thresholds or risk assessment. G6 Audit: Partial. MCP sessions maintain connection state, and tool calls produce structured responses with metadata. However, there is no tamper-evident event log, no hash chain, and no replay guarantee. Audit depends on implementation, not protocol specification. B. A2A v1.0.1 G1 Membership: Partial. Agent Cards declare capabilities and can be registered in directories. The extension mechanism supports new state machines. However, there is no protocol- native admission, invitation, or removal primitive. An agent âexistsâ by publishing an Agent Card; there is no concept of community membership distinct from existence. G2 Deliberation: Absent. A2A supports task delegation and message exchange but not structured argumentation with challenge/response semantics. Messages are task-oriented, not deliberative. G3 Voting: Absent. No voting primitives. The four official extensions (Secure Passport, Timestamp, Traceability, Agent Gateway Protocol) encode none. G4 Dissent: Absent. No dissent semantics in task responses or any extension. G5 Human escalation: Absent. No protocol mechanism for escalating to human authority. Task delegation can target a human-backed agent, but this is routing, not governance escalation with trigger conditions. G6 Audit: Absent. The Traceability extension adds corre- lation IDs for distributed tracing but does not define tamper- evident logs or replay semantics. C. ACP G1 Membership: Partial. ACP defines agent roles within conversations and supports structured multi-party dialogue. However, roles are communication roles (sender, receiver, me- diator), not governance roles (member, moderator, reviewer). There is no admission or removal protocol. G2 Deliberation: Partial. ACPâs negotiation patterns (pro- pose, accept, reject, counter) constitute structured exchange with some challenge/response semantics. However, negotiation is bilateral (between parties with opposing interests), not multilateral deliberation (community reasoning toward shared understanding). The protocol lacks turn-taking governance, relevance enforcement, or synthesis primitives. G3 Voting: Absent. No preference aggregation, quorum, or voting round primitives. G4 Dissent: Absent. Rejected proposals in negotiation are not preserved as dissent in a community record; they are bilateral communication events. G5 Human escalation: Absent. No governance escalation mechanism. G6 Audit: Absent. Message histories exist as conversation state but carry no tamper-evidence or replay guarantees at the protocol level. D. ANP G1 Membership: Absent. ANP routes messages; it does not model community membership. DID-based identity estab- lishes who an agent is, not which community it belongs to. G2 Deliberation: Absent. Routing protocol; no deliberation semantics. G3 Voting: Absent. No voting primitives. G4 Dissent: Absent. No dissent semantics. G5 Human escalation: Absent. No escalation mechanism. G6 Audit: Absent. Message routing may be logged by implementations but the protocol defines no audit primitives. E. ERC-8004 G1 Membership: Partial. The Identity Registry records agent addresses and metadata, functioning as an existence reg- istry. The Reputation Registry gates interactions (agents with low reputation may be excluded). However, there is no for- mal admission protocol, invitation mechanism, or community- scoped membership distinct from global registration. G2 Deliberation: Absent. The specification scopes itself to âdiscover, choose, and interact with agentsâ [5]. No delibera- tion primitives appear in the specification or its listed future work. G3 Voting: Absent. No voting primitives. The Validation Registry records independent attestations but these are unilat- eral validator judgments, not community preference aggrega- tion. G4 Dissent: Absent. No dissent preservation mechanism. G5 Human escalation: Absent. No escalation primitives. G6 Audit: Partial. On-chain transactions are inherently tamper-evident and provide a complete history. However, this is a property of the blockchain substrate, not a governance- specific audit design. The protocol does not define structured event types for decision reconstruction or replay semantics. F. Gap Matrix G. Cross-Protocol Findings Three patterns emerge from the matrix (see also Figure 1): Universal absence (G3, G4, G5). Voting, dissent preserva- tion, and human escalation are absent across all five protocols. No protocolâregardless of its architectural approach (tool- centric, delegation-centric, communication-centric, routing- centric, or trust-centric)âencodes these primitives. This uni- versality suggests the gap reflects a shared design philosophy rather than individual protocol limitations. Partial but insufficient (G1, G2). Membership and de- liberation receive partial treatment: Agent Cards approximate Agent Interoperability Protocol Stack Layer 4: Agent-to-Community (GOVERNANCE) â NO PROTOCOL ENCODES THESE PRIMITIVES G1 Membership G2 Deliberation G3 Voting G4 Dissent G5 Human Esc. G6 Audit/Replay Layer 3: Trust & Reputation ERC-8004 Identity + Reputation + Validation AWS AgentCore Registry + Cedar Auth + Lifecycle Layer 2: Agent-to-Agent (Discovery, Delegation, Communication) A2A v1.0.1 Agent Cards + Task Delegation ACP Negotiation + Roles ANP DID + Graph Routing Layer 1: Agent-to-Tool (Tool Access, Execution) MCP v1.1 Tools + Resources + Prompts + Sessions EXISTING PROTOCOLS THE GAP (This Paper) Fig. 1. Agent interoperability protocol stack. Layers 1â3 (tool access, agent coordination, trust) are addressed by existing protocols. Layer 4 (governance: membership, deliberation, voting, dissent, escalation, audit) is universally absent. capability-based membership; ACP negotiation approximates bilateral deliberation. However, no protocol achieves full sup- port for either dimension. The partial implementations address the coordination aspects of these dimensions (declaring exis- tence, exchanging messages) but not the governance aspects (admission control, structured community deliberation with turn governance). Audit as substrate property (G6). Where audit support exists, it derives from the underlying infrastructure (blockchain immutability for ERC-8004, session state for MCP) rather than from deliberate governance-audit design. No protocol de- fines governance-specific event types, decision-reconstruction semantics, or replay guarantees. V. DISCUSSION A. Extensible vs. Structural Gaps Not all gaps are equal in remediation difficulty. We assess whether each gap could be addressed through the protocolâs existing extension mechanism or requires a new architectural layer. Extensible via A2A: A2Aâs extension mechanism explicitly supports ânew data, requirements, RPC methods, and state machinesâ [7]. Governance primitives (G1âG6) could theo- retically be defined as A2A extensions. The key observation: no one has done so. After 6+ months of A2A being publicly available with an active extension ecosystem, zero governance extensions have been proposed or implemented. Structurally awkward for MCP: MCPâs client-server architecture is tool-centric. Adding community governance to MCP would require agents to participate as both clients and servers simultaneously in a governance contextâa usage pat- tern the protocol was not designed for. Session semantics help TABLE I GOVERNANCE GAP MATRIX: PROTOCOL COVERAGE OF GOVERNANCE DIMENSIONS ProtocolG1 MembershipG2 DeliberationG3 VotingG4 DissentG5 Human Esc.G6 AuditCoverage MCP v1.1AbsentAbsentAbsentAbsentAbsentPartial1/12 A2A v1.0.1PartialAbsentAbsentAbsentAbsentAbsent1/12 ACPPartialPartialAbsentAbsentAbsentAbsent2/12 ANPAbsentAbsentAbsentAbsentAbsentAbsent0/12 ERC-8004PartialAbsentAbsentAbsentAbsentPartial2/12 Any protocolPartialPartialAbsentAbsentAbsentPartialâ Coverage scored as: Supported = 2/2, Partial = 1/2, Absent = 0/2, per dimension per protocol. Maximum possible = 12 (6 dimensionsĂ 2 points each). but do not provide the persistent community state governance requires. Scope-limited for ERC-8004: ERC-8004 could add gover- nance registries (a âDeliberation Registryâ), but its on-chain architecture imposes latency and cost constraints incompati- ble with real-time multi-agent deliberation. The specification explicitly limits scope to discovery and interaction. B. Time-Sensitivity The governance gap is closing, albeit slowly. ERC-8004 absorbed reputationâpreviously identified as a gap [17]â within approximately six months of the gap being articulated. MCPâs specification evolution (from 2024-11-25 to 2025-06- 18 schema) added Elicitation and Sampling but no governance primitives. A2A v1.0.1 added extension infrastructure without governance content. At the observed evolution velocity, we estimate the gov- ernance gap could narrow significantly within 6â12 months through protocol extensions, particularly via A2Aâs exten- sion mechanism. This creates publication urgency for the research community: the window for proposing governance layer designs before de facto standards emerge through ad hoc implementations is narrowing. C. Implications for Deployed Infrastructure The gap extends beyond academic protocols to production infrastructure. AWS Bedrock AgentCore [6] (GA, 15 regions) provides a production agent registry with semantic search, A2A Agent Cards, Cedar-based authorization, lifecycle man- agement, and CloudTrail auditâyet encodes no trust scoring, behavioral reputation, capability verification, or governance primitives. The governance gap is not merely a research find- ing; it affects deployed enterprise systems serving production agent fleets. D. Illustrative Example: What Cannot Be Expressed To make the gap concrete, consider a scenario where an enterprise requires five agents to collectively decide whether a proposed system architecture meets compliance requirements. Listing 1 shows the governance-level messages this interaction requiresânone of which can be expressed in any current protocolâs native semantics. Each message type in Listing 1 maps to a governance dimension (G1âG6). Current protocols can transport these messages as opaque payloads (e.g., via A2A task messages Listing 1. Governance messages required for a collective compliance decision. No current protocol encodes these primitives. // G1: Membership - admit agent to this decision ADMIT agent:security-reviewer TO room:arch-compliance-2026-q3 ROLE: skeptic INVITED_BY: agent:moderator ENDORSEMENTS: [agent:ciso, agent:lead-arch] // G2: Deliberation - structured challenge CHALLENGE claim:c-042 BY agent:security-reviewer TARGETS claim:c-041 (author: agent:arch-proposer) EVIDENCE_REQUIRED: true RATIONALE: "No encryption-at-rest evidence" // G3: Voting - blind preference aggregation VOTE_BLIND claim:c-041 ROUND: 1 VOTER: agent:compliance-officer POSITION: -0.6 // oppose (continuous scale) VISIBILITY: sealed_until_all_cast // G4: Dissent - preserve minority position DISSENT_RECORD claim:c-041 AGENT: agent:security-reviewer POSITION: -0.8 RATIONALE: "Insufficient evidence for..." PRESERVED_IN: decision_record:dr-2026-q3-07 // G5: Human escalation ESCALATE decision:arch-compliance-2026-q3 TRIGGER: mean_confidence < 0.6 ROUTE_TO: human:vp-engineering CONTEXT: [claim:c-041, dissent:d-003] // G6: Audit event EVENT governance:vote_cast ROOM: arch-compliance-2026-q3 ACTOR: agent:compliance-officer PREV_HASH: "a3f8c2..." SIGNATURE: HMAC(actor, payload, prev_hash) or MCP tool calls), but cannot interpret, validate, or enforce their governance semantics. The difference matters: protocol- native governance enables interoperable tooling, standard au- dit formats, and composable governance rules without per- application reimplementation. E. Implications for Protocol Designers Our taxonomy provides a concrete specification for what governance primitives should encode. Protocol designers ex- tending A2A, MCP, or future protocols can use G1âG6 as a requirements checklist, addressing each dimension with protocol-native constructs rather than leaving governance to application-layer reimplementation. F. Limitations This analysis has four limitations. First, we assess speci- fications as of June 2026; protocols evolve rapidly. Second, the âPartialâ classification involves judgment about whether a construct meaningfully addresses a governance dimension. Third, our taxonomy derives from Western organizational theory (Habermas, Robertâs Rules); alternative governance tra- ditions may yield different dimensions. Fourth, we assess what protocols encode, not what can be built on topâdeliberately, as our goal is to identify the architectural gap, but this means practical governance systems can exist without protocol-native support. VI. RELATED WORK Protocol comparison surveys. Ehtesham et al. [18] provide a descriptive comparison of agent communication protocols without governance-specific analysis or a requirements tax- onomy. AgentRFC [19] proposes a six-layer reference stack for agent interoperability but identifies layers rather than gaps within them. Agent governance frameworks. Ruan [12] applies Par- sonsâ AGIL framework to derive a 16-cell institutional ar- chitecture for agent societies, diagnosing governance gaps in MCP/A2A from sociological theory. Our work complements Ruanâs by providing protocol-level (rather than sociological) analysis and producing a concrete gap matrix rather than a theoretical architecture. Bracale Syrnikov et al. [13] introduce governance graphs for preventing LLM collusion, demonstrat- ing that governance mechanisms measurably reduce harmful behavior (collusion from 50% to 5.6%). Their work validates the need for governance; we identify where governance prim- itives are absent. Trust and reputation. Hu and Rong [17] analyze A2A, AP2, and ERC-8004 under six trust models (Brief, Claim, Proof, Stake, Reputation, Constraint). Their taxonomy treats Reputation as a first-class protocol concernânow addressed by ERC-8004. Our taxonomy covers the remaining governance dimensions beyond trust. Multi-agent coordination mechanisms. De Curto et al. [20] propose constitutional multi-agent governance (CMAG) with constraint filtering and penalized-utility opti- mization. Wang et al. [21] introduce conformal social choice for post-hoc deliberation decisions. Gupta et al. [22] imple- ment Ostromâs CPR principles with social learning in LLM agents. These works propose mechanisms for governance; we identify the protocol-level absence that forces each mechanism to reimplement basic governance primitives from scratch. VII. CONCLUSION We presented a systematic analysis of five agent inter- operability protocols against a six-dimension governance re- quirements taxonomy. The gap matrix reveals a clear pattern: current protocols encode coordination (identity, capability, discovery, messaging, reputation) but not governance (mem- bership, deliberation, voting, dissent, escalation, audit). Voting and dissent preservation are universally absent. The gap is not protocol-specificâit reflects a shared design philosophy that treats agents as task workers rather than community participants. This finding establishes that agent community governance is a missing architectural layer, not a missing feature within existing protocols. The community needs governance-native protocol primitivesâmembership admission, structured delib- eration, preference aggregation, dissent retention, and human escalationâencoded at the protocol layer where they can be composed with existing interoperability standards rather than reimplemented ad hoc by every application. The governance gap is narrowing as protocols evolve. We recommend the research community treat the design of governance protocol primitives as an urgent open problem, before the space fragments into incompatible application-layer solutions. AI DECLARATION This research employed Claude (Anthropic) for literature search assistance and draft structuring. All claims, analysis, protocol classifications, and taxonomy derivations were di- rected and verified by the human authors. The gap matrix classifications are based on direct specification analysis by the authors. The authors take full responsibility for all content. REFERENCES [1] Anthropic, âModel context protocol specification,â 2024, schema version 2025-11-25; latest protocol version 2025-06-18. [Online]. Available: https://modelcontextprotocol.io/specification [2] Google and Linux Foundation, âAgent2agent protocol,â 2026, version 1.0.1, May 2026. [Online]. Available: https://a2a-protocol.org/ [3] IBM Research, âAgent communication protocol,â 2025, open-source agent communication framework. [Online]. Available: https://github. com/ibm/agent-communication-protocol [4] ANP Community, âAgent network protocol,â 2025, dID-based agentroutingprotocol.[Online].Available:https://github.com/ agent-network-protocol/ [5] Ethereum Community, âErc-8004: Trustless agents,â 2025, draft, created 2025-08-13. Defines Identity, Reputation, and Validation registries. [Online]. Available: https://eips.ethereum.org/EIPS/eip-8004 [6] Amazon Web Services, âAmazon bedrock agentcore,â 2026, gA, 15 regions. Agent registry, runtime, gateway, evaluations. [Online]. Available: https://aws.amazon.com/bedrock/agentcore/ [7] Linux Foundation, âA2a protocol extensions,â 2026, extension framework documentation, accessed June 2026. [Online]. Available: https://a2a-protocol.org/latest/topics/extensions/ [8] J. Habermas, Between Facts and Norms: Contributions to a Discourse Theory of Law and Democracy. MIT Press, 1996. [9] H. M. Robert et al., Robertâs Rules of Order Newly Revised, 12th ed. PublicAffairs, 2020. [10] E. Ostrom, Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, 1990. [11] C. Sierra, N. R. Jennings, P. Noriega, and S. Parsons, âA framework for argumentation-based negotiation,â in Argumentation in Multi-Agent Systems. Springer, 2004. [12] Y. Ruan, âGovernance by design: A parsonian institutional architecture for internet-wide agent societies,â arXiv preprint arXiv:2604.11337, 2026. [13] A. Bracale Syrnikov et al., âInstitutional ai: Governing llm collusion via public governance graphs,â arXiv preprint arXiv:2601.11369, 2026. [14] Office of the Comptroller of the Currency, âSupervisory guidance on model risk management (sr 11-7),â 2011, oCC Bulletin 2011-12. [15] International Organization for Standardization, âIso/iec 42001:2023 ar- tificial intelligence management system,â 2023. [16] European Parliament and Council, âRegulation (eu) 2024/1689 (artificial intelligence act),â 2024. [17] Y. Hu and W. Rong, âA protocol trust taxonomy for agent interoper- ability,â arXiv preprint arXiv:2511.03434, 2025, aAAI 2026 TrustAgent Workshop. [18] A. Ehtesham et al., âA survey of agent communication protocols,â arXiv preprint arXiv:2505.02279, 2025. [19] AgentRFC Authors, âAgentrfc: A reference framework for agent inter- operability,â arXiv preprint arXiv:2603.23801, 2026. [20] J. de Curto et al., âLlm constitutional multi-agent governance,â arXiv preprint arXiv:2603.13189, 2026. [21] X. Wang et al., âFrom debate to decision: Conformal social choice for safe multi-agent deliberation,â arXiv preprint arXiv:2604.07667, 2026. [22] S. Gupta et al., âThe role of social learning and collective norm formation in llm multi-agent systems,â in Proceedings of AAMAS 2025, 2025.