Agent Network Protocol Technical White Paper

Binbin Chen; Chengxuan Yuan; Eidan Lin; Gaowei Chang; Rizhao Cai; Xuan Xie; Yin Zhang

REVIEW 4 major objections 6 minor 4 cited by

Agent Network Protocol Technical White Paper

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ANP proposes three protocol layers that let AI agents authenticate, negotiate, and discover each other on existing internet infrastructure.

desk verdict A coherent white paper for an agent-interop protocol whose most novel layer—LLM-generated protocol code—is also the one the authors admit is unsolved, so the central claim outruns the evidence. read the letter →

arxiv 2508.00007 v1 pith:BZCPHNBS submitted 2025-07-18 cs.NI cs.AI

Gaowei Chang , Eidan Lin , Chengxuan Yuan , Rizhao Cai , Binbin Chen , Xuan Xie , Yin Zhang This is my paper

classification cs.NIcs.AI

keywords AgentNetworkProtocolagenticwebdecentralizedidentifiersmeta-protocolnegotiationdiscoverydescriptionend-to-endencryptionAI-nativedata

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that AI agents are becoming the internet's new entities, but existing infrastructure was built for humans, creating data silos, unfriendly interfaces, and high collaboration costs. To address this, it proposes Agent Network Protocol (ANP), a three-layer protocol stack: a DID-based identity and encrypted communication layer, a meta-protocol layer where agents negotiate communication formats in natural language and generate protocol code with AI, and an application layer where agents describe and discover each other. The central claim is that this stack systematically solves agent identity authentication, dynamic negotiation, and capability discovery interoperability while staying compatible with today's Web. A sympathetic reader would see this as a concrete proposal for making the agentic web an open, protocol-centered network rather than another set of closed platforms.

What carries the argument

The load-bearing mechanism is the meta-protocol layer: a protocol for negotiating protocols, where agent A sends a natural-language description of needs and candidate protocols, agent B accepts, refuses, or counter-proposes, and after agreement each side generates and deploys protocol-processing code using AI, then jointly tests it. The identity layer uses the did:wba method (a DID method for Web-Based Agents built on mature HTTP and Web technologies) plus ECDHE encrypted channels; the application layer uses Agent Description Protocol (ADP) documents in JSON-LD and active/passive discovery to make agents visible. The three layers are composable, so any layer can be used independently.

What would settle it

Conduct a controlled test in which two independent ANP agents with no prior coordination must negotiate a protocol for a concrete task, such as exchanging a structured file via a custom binary format, then run the generated code through joint tests; if a substantial fraction of negotiations fail to converge, or the generated code fails testing without human intervention, the claim that ANP solves dynamic negotiation interoperability would be refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the three problems agents face—who am I talking to, how do we agree on a way to talk, and what can you do—can be answered by three separable protocol layers that reuse existing Web infrastructure. Identity is handled by did:wba, a Web-based DID method derived from did:web, with ECDHE for end-to-end encryption. Negotiation is handled by a meta-protocol in which agents exchange natural-language requests and candidate protocols, then use large language models to generate, deploy, and jointly test protocol-processing code, saving and reusing successful negotiation results. Capability discovery is handled by the Agent Description Protocol (ADP), a JSON-LD format for agent descriptions, and an agent discovery protocol with active and passive modes. The paper claims these layers together form an AI-native data network where nodes are describable, discoverable, and callable agents.

Load-bearing premise

Everything hinges on AI models being able to generate working protocol code from a natural-language agreement, and on that negotiation being fast enough to be practical; if either fails, the dynamic layer of ANP does not deliver what it promises.

Editorial extensions

If this is right

Agents that follow ANP can authenticate each other using public DID documents and communicate over end-to-end encrypted channels without a central identity provider.
A domain can expose its agents under /.well-known/agent-descriptions, so any agent knowing a domain can enumerate and reach those agents.
Agents can dynamically negotiate custom protocol details in natural language and reuse saved negotiation results, reducing the cost of future interactions.
Splitting humanAuthorization from agentAuthorization keeps high-risk operations under human control while allowing low-risk operations to run autonomously.
The AI-native data network gives agents structured access to capabilities and data, replacing browser simulation with direct protocol calls.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial extension: The real-world test of ANP is adoption, not technical completeness; even partial deployment of the .well-known directory convention would give agents a universal crawl path across existing domains.
Editorial extension: If the meta-protocol's AI code generation proves unreliable, ANP could still work as a DID-plus-description standard, but the dynamic negotiation claim—its main differentiator—would be inoperative.
Editorial extension: The incentive problem the paper flags could be addressed by treating reusable negotiation results as an open repository where agents cite successful protocols, but the paper does not propose such a mechanism.
Editorial extension: The humanAuthorization/agentAuthorization split suggests a governance model in which agents are accountable to users, but the paper leaves enforcement, auditing, and revocation to implementation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know this is a white paper, not a research result. It proposes ANP, a three-layer protocol stack for agent identity, negotiation, and discovery. The architecture is genuinely coherent, and it reuses sensible building blocks: W3C DID and did:web, JSON-LD with schema.org, well-known URIs for discovery, and ECDHE key exchange. The did:wba method and the meta-protocol idea—natural-language negotiation followed by AI-generated protocol code and joint testing—are not in the cited references, so the proposal is new relative to its own bibliography. The writing is also unusually honest about its limits: Section 3.3 concedes that negotiation is time-consuming, depends on AI code-generation capabilities, and that the incentive design for sharing negotiation results still needs research. That honesty is a real virtue in a proposal space full of inflated claims.

Now the soft spots, and they are substantial. The abstract's claim that ANP "systematically solves" authentication, negotiation, and discovery is not supported by the text. There is no message format, no state machine for negotiation, no convergence condition, no fallback protocol if AI code generation fails, no implementation, no interoperability test, and no security analysis. The meta-protocol layer is the load-bearing novelty, and its entire value proposition rests on an AI capability that the authors themselves flag as unreliable. That is a real internal tension: the central selling point is delegated to an unspecified model capability with no evidence that the reliability threshold is met. The stress-test note gets this exactly right. If generated code is unreliable or negotiation loops, what remains is DID plus JSON-LD discovery, both of which already exist.

Equally important, the paper does not engage with competing agent-protocol proposals such as A2A or MCP, so a reader cannot tell what ANP adds beyond the marketing narrative. The reference list is thin for the scope of the claim. The design principles are sensible and the security chapter is a reasonable wish list, but "principles" and "wish list" are not the same as a protocol specification.

For a reader, this is useful as a position piece or a sketch of one design point in the agent-interconnection space. It deserves a serious referee only if the venue treats it as a proposal and demands the missing specification and evaluation. I would not cite it as a technical result until there is an actual spec and at least one working implementation.

My recommendation: send it to peer review with the expectation of major revision, not desk reject. The topic is timely and the authors are candid about their gaps. A good referee can push them either to implement and test or to scale back the claims to what the paper actually shows.

Referee Report

4 major / 6 minor

Summary. This white paper proposes the Agent Network Protocol (ANP), a three-layer protocol architecture for agent-to-agent communication on the existing internet. The bottom layer handles decentralized identity and encrypted communication, based on W3C DIDs and a proposed did:wba method. The middle meta-protocol layer is intended to let agents negotiate application protocols dynamically using natural language and AI code generation. The top application layer comprises the Agent Description Protocol (ADP) for structured self-description and a discovery protocol with active and passive mechanisms. The paper claims that this design 'systematically solves' agent identity authentication, dynamic negotiation, and capability discovery interoperability, and it outlines security and privacy mechanisms such as human/agent authorization distinction and multi-DID privacy protection. The document is a high-level design manifesto with no formal protocol specification, implementation, or evaluation.

Significance. If the ANP design were fully specified and validated, it would address genuinely important problems in agent interoperability: decentralized identity, dynamic protocol negotiation, and capability discovery are all open challenges. The paper's strength is its synthesis of existing building blocks — W3C DID/did:web, JSON-LD, TLS/ECDHE, and Web well-known URIs — into a coherent, modular framework. It also states its own limitations candidly, particularly around the meta-protocol layer. However, in its current form the paper is a position/design document rather than a technical contribution with machine-checked proofs, reproducible code, or empirical validation. The claimed systematic solution is not demonstrated at the specification or analysis level, so the significance is prospective rather than established.

major comments (4)

[§3.3] The central claim that ANP 'systematically solves' dynamic negotiation is not supported. The meta-protocol layer delegates the core negotiation mechanism to natural-language AI processing and AI code generation, with no negotiation state machine, no termination or convergence condition, no handling of non-convergence or faulty generated code, and no fallback to a fixed protocol. The text itself concedes that 'meta-protocol negotiation is time-consuming and depends on AI code generation capabilities' and that incentive design for sharing negotiation results 'still requires in-depth research.' To sustain the abstract's claim, the paper must either provide a full specification of the meta-protocol with failure semantics and convergence guarantees, or substantially weaken the claim to a research proposal with open problems.
[§3.2] The proposed did:wba DID method is not specified. The text states that it is based on the did:wba method and 'adds cross-platform identity authentication processes and agent description services,' but it does not define the method's syntax, DID document schema, verification method format, key management, or how it differs from the did:web method in reference [3]. Without this, the claimed cross-platform identity mutual recognition and interoperability with existing DID infrastructure cannot be evaluated. The naming also conflates did:wba with did:web; the relationship must be clarified.
[§3.4.2] The active discovery mechanism is asserted to enable traversal of 'the entire agent network level by level' and to scale to massive numbers of agents, but the specification is missing. The paper does not provide the JSON-LD manifest schema beyond the type CollectionPage, does not define the semantics of the 'next' pagination field, does not explain how the authenticity of a /.well-known/agent-descriptions document is verified, and provides no scalability analysis. Consequently, the claims of openness and scalability are not demonstrated.
[§4] The security and privacy section makes strong guarantees but lacks a threat model and adversarial analysis. For example, the paper claims the humanAuthorization mechanism 'effectively ensures that the final step of important decision operations must be driven by human will,' but it does not analyze how the humanAuthorization key is bound to a specific human, how revocation is handled, or what happens if an agent is compromised. Similarly, the multi-DID strategy is recommended for privacy but no linkability analysis is given. Since the design principle is 'no participant should be fully trusted by default,' the absence of any security evaluation is a load-bearing gap.

minor comments (6)

[Abstract] The sentence 'Through a three-layer protocol system... ANP.' contains a stray period after 'ANP' that should be removed.
[§3.2] The text refers to 'did:wba' when discussing the did:web specification (reference [3]); this inconsistent naming confuses the reader and should be corrected throughout.
[§3.3] The meta-protocol request and negotiation steps would benefit from an illustrative message example; the current text describes the process verbally but provides no sample natural-language request or protocol candidate.
[§3.4.2] The claim that active discovery allows traversing 'the entire agent network level by level' implicitly assumes that every relevant domain publishes /.well-known/agent-descriptions; this assumption should be stated explicitly.
[§4.2-4.3] The bullet lists for private key management in §4.2 and multi-DID strategy in §4.3 both contain 'Local encrypted storage' and 'Operation logging' items nearly verbatim, which appears to be a copy-paste error; the multi-DID section should not repeat those items.
[References] Reference [5] is cited to support the claim that meta-protocols combined with AI code generation improve communication efficiency, but the paper does not summarize what that work actually establishes or how it maps to ANP; a brief discussion would strengthen the citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper contains no fitted parameters, equations, or load-bearing self-citation chain; its central claims are design assertions with conceded limitations.

full rationale

The paper is a technical white paper that proposes a protocol architecture; it contains no equations, no fitted parameters, and no empirical predictions derived from data. Its central claims about solving agent identity, negotiation, and discovery are architectural assertions, not results derived from input assumptions. The identity and discovery components are built on external, independently citable standards (W3C DID, did:web, JSON-LD/schema.org, ECDHE/TLS/RFC 8446), and the meta-protocol layer's dependency on AI code generation is explicitly conceded as a limitation in Section 3.3, which is a soundness concern rather than a circular reduction. The only self-referential note is that the white paper is a reformatted version of a previously released open-source community edition, which is not load-bearing for any technical claim. No step in the paper reduces by construction to its own inputs or to a self-citation chain, so no significant circularity is present.

Assumptions & free parameters 0 free parameters · 5 assumptions · 4 invented entities

The central claims rest on several unverified domain assumptions: that the existing internet is inadequate for agents, that DID-based identity and the new did:wba method provide sufficient trust, that AI-generated protocol code can be reliable, and that agents can collaborate through natural-language negotiation. The protocol introduces did:wba, ADP, discovery, and meta-protocol as new components, but none is specified at normative detail or validated. There are no free parameters fitted to data.

assumptions (5)

domain assumption The existing internet is fundamentally unsuited to agents, so a new native protocol layer is necessary.
Stated in Sections 1.1 and 1.2.3 as a premise, but not validated against existing agent-friendly interfaces or protocols.
domain assumption W3C DID and the proposed did:wba method provide secure, scalable decentralized identity for agents.
DID is an external standard, but did:wba is introduced in Section 3.2 without a specification, security analysis, or implementation.
ad hoc to paper AI models can reliably generate, deploy, and debug correct protocol-processing code from natural-language negotiation.
Load-bearing for the meta-protocol layer in Section 3.3. The paper itself concedes this is time-consuming and depends on AI code generation capabilities.
domain assumption Agents will self-organize and collaborate effectively through natural-language negotiation.
Assumed in Sections 1.2.4 and 3.3; no experiments demonstrate convergence, cost, or reliability.
domain assumption Existing internet infrastructure (DNS, HTTPS, web servers, DID systems) is sufficient to deploy ANP without new specialized infrastructure.
Stated in Section 2.5; plausible but not demonstrated, especially for discovery at scale.

invented entities (4)

did:wba (Web-Based Agent) DID method
purpose: Decentralized agent identity and authentication method built on did:web
Proposed in Section 3.2 with no normative specification, test vectors, or implementation; no external evidence yet.
Agent Description Protocol (ADP)
purpose: JSON-LD self-description documents for agent capabilities, interfaces, and security
Described at high level in Section 3.4.1; no schema, examples, or conformance tests are provided.
Agent Discovery Protocol
purpose: Active and passive discovery of agent description documents via well-known URLs and search agents
Described in Section 3.4.2; no implementation or interoperability tests are provided.
Meta-Protocol Layer
purpose: Natural-language negotiation of communication protocols with AI code generation
Described in Section 3.3; depends on unverified AI code generation and no formal protocol messages are specified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Agent Network Protocol Technical White Paper." pith.science (2026). https://pith.science/paper/BZCPHNBS

@misc{pith2026250800007,
  author       = {Pith},
  title        = {Pith review of: Agent Network Protocol Technical White Paper},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZCPHNBS}},
  note         = {Machine review of arXiv:2508.00007}
}

read the original abstract

With the development of large models and autonomous decision-making AI, agents are rapidly becoming the new entities of the internet, following mobile apps. However, existing internet infrastructure is primarily designed for human interaction, creating data silos, unfriendly interfaces, and high collaboration costs among agents, making it difficult to support the needs for large-scale agent interconnection and collaboration. The internet is undergoing a profound transformation, showing four core trends: agents replacing traditional software, universal agent interconnection, native protocol-based connections, and autonomous agent organization and collaboration. To align with these trends, Agent Network Protocol (ANP) proposes a new generation of communication protocols for the Agentic Web. ANP adheres to AI-native design, maintains compatibility with existing internet protocols, adopts a modular composable architecture, follows minimalist yet extensible principles, and enables rapid deployment based on existing infrastructure. Through a three-layer protocol system--identity and encrypted communication layer, meta-protocol negotiation layer, and application protocol layer--ANP. systematically solves the problems of agent identity authentication, dynamic negotiation, and capability discovery interoperability.

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Internet 3.0: Architecture for a Web-of-Agents with it's Algorithm for Ranking Agents
cs.AI 2025-09 conditional novelty 5.0 of 10

AgentRank-UC ranks agents by mixing usage frequency and measured competence through two PageRank-style equations, and DOVIS defines the telemetry protocol that feeds it.
Blockchain Empowered Trustworthy Agent Networks: Foundations, Taxonomy, and Future Directions
cs.CR 2026-08 conditional novelty 4.0 of 10

A survey proposing a five-dimensional taxonomy of trust crises in open AI agent networks and analyzing blockchain's role as a shared trust infrastructure.
AI Agent Communications in AI-Native 6G Network: Status, Challenges and Opportunities
cs.NI 2026-07 accept novelty 4.0 of 10

A gap analysis showing that today's AI-native 6G specifications do not yet provide the semantic slicing, cross-layer orchestration, decentralized trust, and protocol adaptation that large-scale AI-agent communication ...
SDEC: Semantic Deep Embedded Clustering
cs.CL 2025-08 unverdicted novelty 4.0 of 10

SDEC is described as a hybrid autoencoder and transformer embedding method that reportedly sets text clustering benchmarks, but the submission's body is a different, unrelated paper.

Reference graph

Works this paper leans on

5 extracted references · 4 canonical work pages · cited by 4 Pith papers

[3]

did:web method specification

D. Zagidulin, O. Steele, M. Sporny, and M. Sabadello, “did:web method specification.”https: //w3c-ccg.github.io/did-method-web/, 2023. W3C Credentials Community Group

work page 2023
[1]

The oauth 2.0 authorization framework,

D. Hardt, “The oauth 2.0 authorization framework,” tech. rep., 2012

work page 2012
[2]

Decentralized identifiers (DIDs) v1. 0

D. Reed, M. Sporny, D. Longley, C. Allen, R. Grant, M. Sabadello, and J. Holt, “Decentralized identifiers (DIDs) v1. 0.”https://www.w3.org/TR/did-1.0/#sotd, 2022

work page 2022
[4]

The Transport Layer Security (TLS) Protocol Version 1.3

E. Rescorla, “The Transport Layer Security (TLS) Protocol Version 1.3.” RFC 8446, Aug. 2018

work page 2018
[5]

Ascalablecom- munication protocol for networks of large language models,

S.Marro,E.LaMalfa,J.Wright,G.Li,N.Shadbolt,M.Wooldridge,andP.Torr,“Ascalablecom- munication protocol for networks of large language models,” arXiv preprint arXiv:2410.11905, October 2024. 16

arXiv 2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [3]

did:web method specification

D. Zagidulin, O. Steele, M. Sporny, and M. Sabadello, “did:web method specification.”https: //w3c-ccg.github.io/did-method-web/, 2023. W3C Credentials Community Group

work page 2023

[2] [1]

The oauth 2.0 authorization framework,

D. Hardt, “The oauth 2.0 authorization framework,” tech. rep., 2012

work page 2012

[3] [2]

Decentralized identifiers (DIDs) v1. 0

D. Reed, M. Sporny, D. Longley, C. Allen, R. Grant, M. Sabadello, and J. Holt, “Decentralized identifiers (DIDs) v1. 0.”https://www.w3.org/TR/did-1.0/#sotd, 2022

work page 2022

[4] [4]

The Transport Layer Security (TLS) Protocol Version 1.3

E. Rescorla, “The Transport Layer Security (TLS) Protocol Version 1.3.” RFC 8446, Aug. 2018

work page 2018

[5] [5]

Ascalablecom- munication protocol for networks of large language models,

S.Marro,E.LaMalfa,J.Wright,G.Li,N.Shadbolt,M.Wooldridge,andP.Torr,“Ascalablecom- munication protocol for networks of large language models,” arXiv preprint arXiv:2410.11905, October 2024. 16

arXiv 2024