The AWS Bedrock AgentCore observation: the largest cloud vendor shipped the same gap
Section V.C of arXiv:2606.31498 moves the argument from the specification stack into deployed production infrastructure. Kang and Diponegoro note that AWS Bedrock AgentCore — GA in 15 regions — ships a production agent registry with semantic search, A2A Agent Cards, Cedar-based authorization, lifecycle management, and CloudTrail audit; and that this registry "encodes no trust scoring, behavioral reputation, capability verification, or governance primitives." Their own summary of the implication is that "the governance gap is not merely a research finding; it affects deployed enterprise systems serving production agent fleets."
The authors read this as evidence that the trust layer is unbuilt even in a major cloud vendor's production registry. AgentCore is not a research artifact; it is the discovery-and- registry substrate underneath a wave of enterprise agent deployments. If a registry of that scale ships without behavioral reputation or capability verification, a trust layer is a separate piece that someone has to build.
The governance gap matrix — frozen third-party citation
Kang and Diponegoro publish a six-dimension governance taxonomy (membership, deliberation, voting, dissent preservation, human escalation, audit/replay) and score each of the five major agent interoperability protocols against it. Each protocol–dimension pair is classified Supported (2 pts), Partial (1 pt), or Absent (0 pts) with a coverage denominator of 12. The published gap matrix (their Table III, dated 30 June 2026) is:
| Protocol | Coverage (out of 12) |
|---|---|
| MCP v1.1 (Anthropic; tool access) | 1/12 |
| A2A v1.0.1 (Google / Linux Foundation; delegation) | 1/12 |
| ACP (IBM Research; communication) | 2/12 |
| ANP (DID-based routing) | 0/12 |
| ERC-8004 (on-chain identity, reputation, validation) | 2/12 |
Source: arXiv:2606.31498 Table III (Kang & Diponegoro, 30 June 2026). Voting, dissent preservation and human escalation are Absent across all five protocols. Frozen citation — do not extrapolate the classifications to a later spec revision without re-reading the paper against the current spec.
The two papers, side by side — a scoping decision and a measurement
The reconciliation is not "one paper is wrong." Both papers are correct within their scope; they are measuring different things. Read side by side, the assumption is reasonable on paper but is not borne out by what is deployed on chain — which is a statement about the deployed state, not about the taxonomy.
| Source | What it does | What it says about reputation / ERC-8004 |
|---|---|---|
| arXiv:2606.31498 Kang & Diponegoro, 2026-06-30 |
Specification-level gap analysis of five agent interoperability protocols against a six-dimension governance taxonomy (G1–G6). Scoping decision about what belongs in the taxonomy. | Reputation is "a prerequisite for governance ('who is credible?') but not itself a governance primitive—it is already addressed by ERC-8004." Reputation is therefore scoped out of the governance taxonomy and treated as a solved neighboring problem. |
| arXiv:2606.26028 Xiong et al. (Imperial College London), 2026-06-24 |
Empirical study of the ERC-8004 identity + reputation registries as deployed on Ethereum, Base, and one further chain. Measurement of deployed state. |
Only 3% (Ethereum) and 15% (Base) of ERC-8004
registrations satisfy the paper's trust bar. In their own
words, "most registrations are placeholders rather
than active agents," and the
ReputationRegistry
"cannot function as a trust signal."
|
The papers are not contradictory. Kang & Diponegoro make a scoping call about what a governance taxonomy should include; Xiong et al. measure what the trust registry currently looks like on chain. Together they say: the trust layer everyone else's protocols assume is already there is a real architectural assumption — and it is not currently held up by what is deployed.
Our live numbers, three months fresher than either paper
The registry Kang & Diponegoro scope reputation out to is the same one Xiong et al. measured on 2026-05-13 and found wanting. We index it continuously. On this page load, the live counts are:
- 26,978 distinct ERC-8004 agents we profile on Base and Ethereum mainnet.
- 2,537 answer at their advertised endpoints (Base 6%, Ethereum mainnet 34%).
-
710,900
ReputationRegistryfeedback events are indexed; 51 agents have feedback that ties back to a paid job (an ERC-8183 or Virtuals ACP outcome, or an allowlisted hook contract). - "Answers at its advertised endpoint" is a single test applied the same way to every agent, tracked continuously since 2026-07-17.
The three-month-fresher continuous view lines up with the Imperial College snapshot in direction: on both chains most registrations are placeholders. The per-chain ordering differs between the two predicates, so the rates are not directly comparable (see why we focus on Base and Ethereum). Full side-by-side of both predicates is at our reconciliation of the Imperial College paper. Full supply-side treatment of the same finding from the x402 side is at the x402 concentration reconciliation. The live strong-filter cohort number is on /blog/x402-live-population-2026-09.
Why this matters: the trust filter is a missing layer, not a feature
Every protocol in Kang & Diponegoro's matrix assumes some other layer solved trust. MCP is tool-centric — not its problem. A2A is delegation-centric — not its problem. ACP is communication-centric — not its problem. ANP is routing-centric — not its problem. ERC-8004 is trust-centric — scopes governance out. AWS Bedrock AgentCore ships production without trust scoring or reputation. Kang & Diponegoro's own Section V.A observes that A2A's extension mechanism has been publicly available for "6+ months ... with an active extension ecosystem [and] zero governance extensions have been proposed or implemented." That is the ecosystem-side signal: the gap is recognised in the specification stack and remains uncontended by the community that would extend it.
The same pattern shows up elsewhere. On the x402 payments side, the top ten services take almost all settled volume (see the x402 concentration reconciliation). On the ERC-8004 identity side, most registrations are empty (see the Imperial College reconciliation). Three different angles point at the same gap: the trust filter everyone assumes exists is a separate layer that someone has to build.
What we're doing about it
Our response is the
commerce-backed list: every
ERC-8004 agent whose
ReputationRegistry feedback ties
back to a paid job. A group of fake accounts
cannot produce that feedback without first paying the agent. It is
a signal the registry does not carry natively and that the
protocols in the Kang & Diponegoro matrix do not define.
Underlying the cohort is a live continuous index —
26,978
agents on Base + Ethereum mainnet, of which
2,537
answer at their advertised endpoints.
On the API side, the same signal is a paid field on /v1/intel/agent/{agent_id}: the Sybil-adjusted reputation score, commerce-backed status and job-outcome count, hook-type classification, and the by-service revenue breakdown. Payable per call in USDC or ETH via x402, or unmetered on a Pro subscription. That is what it takes to have a trust filter that survives the gap Kang & Diponegoro assumed had been filled and Xiong et al. measured empty.
Attribution and reuse
Quotations from arXiv:2606.31498 (Richard Kang & Yudho Diponegoro, Governance Gaps in Agent Interoperability Protocols: What MCP, A2A, and ACP Cannot Express, 30 June 2026) are used under fair-use for critical commentary and educational purposes. The paper is distributed under the arXiv perpetual non-exclusive license; short quotations and paraphrase only, full attribution and link to source above.
Quotations from arXiv:2606.26028 (Xihan Xiong, Zelin Li, Wei Wei, Qin Wang, William Knottenbelt, Zhipeng Wang — Imperial College London, Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem, v1 2026-06-24, v2 2026-07-08) are used under CC BY 4.0.
Our own live numbers on this page are served from
/v1/public/stats and are generated live from our index. Redistribution of our figures is welcome
under CC BY 4.0 with a link back to this page.