Blog · arXiv Analysis · Last reviewed June 25, 2026

The Agent Standard Becomes the Governance Graph

Yutian Wang and Luyao Zhang's June 2026 arXiv paper Agentic Analysis for Agentic Infrastructure compares the governance records of ERC-8004 and Google A2A. Its useful lesson is that agent interoperability standards do not only define messages, registries, and handshakes. They also leave a graph of who participated, which issues were made visible, and where authority actually sat.

The Standard Has a Room

Wang and Zhang's paper, arXiv:2606.26203 [cs.AI], was submitted on June 24, 2026. The arXiv record lists the title as Agentic Analysis for Agentic Infrastructure: An LLM-Powered Pipeline for Comparative Governance of DAO and Corporate AI Protocols and the authors as Yutian Wang and Luyao Zhang.

The paper studies two agent interoperability standards that already have separate pages on this site: ERC-8004, an Ethereum agent-trust proposal, and Google A2A, a protocol for agent-to-agent communication now organized under the A2A project. This essay is not another explainer of either protocol. It is about the rooms where protocol text is made.

That matters because standards can present themselves as neutral plumbing. In agent systems, plumbing decides who can discover whom, how delegation is represented, what metadata must be exposed, what authority can be verified, and which errors become auditable. A protocol standard is therefore a technical document and a constitutional artifact.

What the Graph Measures

The authors build an LLM-assisted pipeline for governance discourse analysis. They collect public participation records, annotate argumentative function, stance, and stakeholder affiliation, then combine topic modeling with co-participation, discourse-network, and socio-semantic network analysis. The main corpus contains 4,323 retained records: 142 for ERC-8004 and 4,181 for Google A2A.

The method is useful because participation count alone is a weak governance signal. A standard can have many comments and still be shaped by a small working core. A decentralized process can be open and still be socially concentrated. A corporate process can be public on GitHub while decisive coordination happens in steering calls, partner channels, or design reviews that are absent from the public trace.

The paper's question is therefore not "which protocol is better?" It asks how governance form shapes topic selection, participation structure, consensus, conflict, and actor-topic labor. That is exactly the right layer for agent infrastructure: before agents coordinate through a standard, humans and institutions coordinate over what the standard will recognize.

Two Forms of Rule

In the paper's framing, ERC-8004 and A2A address a similar technical domain but differ sharply in governance architecture. ERC-8004 proceeds through Ethereum-style public discussion, rough consensus, editor review, and permissionless implementation. A2A began as a Google-led protocol and, according to the A2A governance materials cited by the paper, runs through a technical steering committee during its startup phase.

The contrast is not purity versus corruption. The paper is more interesting than that. ERC-8004's public record concentrates on constitutive questions: trust, security mechanisms, protocol principles, reputation, verification, and what the standard should mean. A2A's record is broader and more execution-oriented, with attention spread across documentation, examples, transport, versioning, collaboration, and project process.

Those differences are governance outputs. A standard formed around on-chain trustless-agent registries will naturally draw one set of anxieties. A standard formed around enterprise agent interoperability will draw another. The Spiralist point is that a protocol's politics are not downstream from implementation. They are visible in the issue labels, forum replies, pull requests, steering rules, and themes that survive into the public archive.

The Elite Does Not Disappear

One of the paper's most useful findings is that openness does not erase concentration. It reports high participation inequality in both cases, with degree Gini values of 0.804 for ERC-8004 and 0.779 for A2A. The top contributors, whether volunteer specialists, protocol insiders, or corporate representatives, still do much of the visible work.

That should discipline how we talk about open agent standards. Open discussion can improve observability, invite critique, and preserve a public record. It does not automatically distribute influence evenly. Corporate governance can produce shipping discipline and implementation support. It does not automatically make its decisive negotiations visible.

The paper also finds that discourse congruence is denser in the ERC-8004 setting, while A2A produces more absolute conflict volume in the larger public record. In plain terms: the permissionless case shows tighter public agreement within its communities, while the corporate-led case leaves a larger, more fragmented engineering trace.

The Observatory Limit

The authors are careful about limits, and this is where the paper earns trust. The ERC-8004 main corpus is much smaller than the A2A corpus. Public GitHub and forum traces do not capture every meeting, private chat, partner negotiation, or steering call. The paper's appendix says A2A's technical steering committee meetings, internal Google design reviews, and partner negotiations may occur outside the public repository, so public fragmentation may understate private coordination.

LLM-assisted annotation is also not neutral magic. The authors run robustness checks with multiple models, but the labels still depend on codebooks, prompts, model behavior, and human review. The page should therefore be read as evidence about public governance traces, not as a final moral ranking of A2A, ERC-8004, DAOs, corporations, or open source.

Governance Standard

An agent standard should publish a governance graph beside its specification. The graph should record decision rights, eligible voters, maintainers, editors, steering members, affiliation labels, public and private decision channels, issue and pull-request counts, abstentions, rejected proposals, conflict-resolution rules, and version milestones.

For agent infrastructure, this is not decorative transparency. The standard will shape identity, delegation, audit, payments, discovery, and trust among systems that act for people and organizations. If the protocol becomes part of machine bureaucracy, the public should be able to inspect not only the syntax agents use, but the social machinery that decided the syntax.

Sources

Return to Blog