Pith. sign in

REVIEW 3 major objections 4 minor 33 references

This paper argues that AI tool discovery can be reduced to DNS name resolution, cutting per-query search space by 95.26% while matching flat retrieval accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 15:54 UTC pith:YCATAJ3V

load-bearing objection ToolDNS has a genuinely interesting idea, but its central compatibility claim fails as written; the evaluation metrics are too weak to support the accuracy claims. the 3 major comments →

arxiv 2607.18242 v1 pith:YCATAJ3V submitted 2026-04-19 cs.AI cs.MAcs.NI

AI Tool Discovery at Scale: All You Need is DNS

Enhao Chen , Yulin Shao This is my paper
classification cs.AI cs.MAcs.NI
keywords AI tool discoveryDNSsemantic discoveryEDNS0hierarchical namespaceservice discoverylarge language modelsdecentralized governance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that AI-agent tool discovery—finding which function, API, or skill an agent should invoke—can be reduced to a naming problem and solved with the existing DNS infrastructure rather than a new centralized registry. The proposed system, ToolDNS, encodes functional intent and organizational trust into a hierarchical namespace under a .tools domain, and turns each discovery query into a sequence of lightweight DNS name resolutions. Three standards-compliant extensions carry the mechanism: partially unfolded domain names act as a search cursor, EDNS0 options transmit the natural-language intent, and logical subdomains let independent organizations govern their own tool branches. On a released benchmark of 33,688 real tools across four protocols, the two-level hierarchy cuts the per-query search space by 95.26% while matching flat vector-retrieval accuracy, and its UDP-native design lowers network overhead. If correct, this would give agent ecosystems a scalable, decentralized discovery fabric without new global infrastructure.

Core claim

The central claim is that service discovery for AI agents is better modeled as hierarchical name resolution than as flat semantic matching. ToolDNS maps each tool to a node in a semantic namespace, and resolves an agent's intent by iteratively narrowing the namespace: at each level an authoritative server uses the EDNS0-carried intent to rank its child subdomains and returns the top-K as NS records. Partially unfolded names encode the current search position in the query name itself, so no per-query state is held by the resolver. This replaces the O(N) similarity scan of vector retrieval with O(beta*h + S) work, which is O(log N) on a balanced tree. Empirically, with only two hierarchy level

What carries the argument

The load-bearing object is the semantic namespace itself: a tree under a reserved .tools top-level domain where each label is a functional category and each leaf hosts tool SRV records. Three protocol-compliant mechanisms make it work: partially unfolded domain names, where an underscore before the domain acts as a cursor that the resolver replaces with the next matched subdomain; EDNS0 option payloads that carry the agent's natural-language intent and a K parameter to every authoritative server; and logical subdomains that decouple a functional prefix (weather.tools) from organizational identifiers (hku.weather.tools), enabling multi-tenant governance. The authoritative servers do the seman

Load-bearing premise

The scheme collapses if ordinary, unmodified recursive resolvers will not forward the custom EDNS0 payload and follow the NS referrals for partially unfolded names, because the claimed zero-modification deployment depends on that behavior.

What would settle it

Point a stock, unmodified recursive resolver at a .tools test zone, send it a partially unfolded SRV query with an EDNS0 intent payload (e.g., _mcp._tcp._weather.tools), and observe whether it follows the returned NS referral and rewrites the cursor. If the resolver returns NXDOMAIN or refuses to follow the referral, the protocol's compatibility premise is empirically false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Tool discovery can scale to millions of tools with per-query cost that grows logarithmically instead of linearly, provided the taxonomy is balanced.
  • New or updated tools become discoverable immediately, since matching happens at query time against local zone data and no global index needs rebuilding.
  • A single discovery interface covers tools that use different invocation protocols, because SRV service and protocol labels encode the protocol in the name.
  • Agents can enforce trust and security policies by resolving only within organizational subdomains, without a central trust anchor.
  • The hierarchical decomposition itself improves LLM ranking accuracy by reducing attention dilution, beyond the efficiency gain.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the same intent-to-namespace encoding could apply to other agent-facing resources—models, datasets, or even other agents—turning any registration problem into a DNS zone.
  • Editorial extension: the claimed latency advantage depends on staying in plain UDP; if EDNS0 payloads grow or encrypted transports like DoT/DoH become mandatory, the messaging overhead may approach that of HTTP, so the comparison should be re-run under encrypted transports.
  • Editorial extension: the search-space reduction is only as good as the taxonomy; a stress test with misclassified or overlapping categories, or with tools that legitimately span multiple branches, would show how robust the pruning is.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes ToolDNS, a DNS-based hierarchical tool discovery framework. Functional intent is encoded into partially unfolded domain names and EDNS0 payloads; LLM-based semantic scoring at authoritative servers prunes candidate subdomains level by level. The paper claims O(log N) per-query complexity, a 95.26% search-space reduction, retrieval accuracy comparable to state-of-the-art flat retrieval, and compatibility with unmodified DNS resolvers. It also contributes a 33,688-tool benchmark spanning MCP, A2A, RESTful, and Skill protocols. Experiments report category-level hit rates, search-space counts, and network-traffic comparisons against HTTP-based baselines.

Significance. The underlying vision of reusing DNS for agent tool discovery is thought-provoking, and the released multi-protocol dataset is a potentially useful community resource. The logical-subdomain idea for delegated trust is also interesting. However, the central deployment claim—that ToolDNS runs on unmodified recursive resolvers—is contradicted by the paper's own algorithm and compatibility discussion. In addition, the evaluation is substantially self-referential: the taxonomy, the ground-truth queries, and the semantic scorers all come from the same LLM family, and the hit-rate metric counts any tool in the same leaf sub-category as a hit. As presented, the paper does not support its headline claims of matching state-of-the-art retrieval accuracy and zero-modification deployment.

major comments (3)
  1. [§3.3.2, §3.4 Algorithm 1, §4.2] The claim that ToolDNS works with unmodified recursive resolvers is contradicted by Algorithm 1 and by the paper's own Section 4.2. Algorithm 1 requires the resolver to select a subdomain from returned NS records and rewrite the QNAME by replacing the cursor label (line 14). Standard iterative resolvers follow referrals only when the current QNAME is a descendant of the delegated zone (RFC 1034). The initial QNAME `_mcp._tcp._tools.` is not a descendant of `weather.tools`, and the derived name `_mcp._tcp._weather.tools.` is produced by nonstandard client/resolver logic. Section 4.2 concedes that unmodified resolvers 'will not perform semantic pruning, so the query will likely fail to resolve.' Thus the zero-modification, no-new-infrastructure contribution is not supported; a custom resolver or custom iterative client is required. This is a load-bearing issue, not a presentation detail.
  2. [§5.1, §5.2, Definition 2] The accuracy evaluation is circular and the metric is too lenient. The hierarchical taxonomy used by ToolDNS's semantic pruning is constructed by Qwen/DeepSeek in Steps 3–6 of the data pipeline, and the ground-truth query–tool pairs are generated by Qwen in Step 7. The same model family is then used as the semantic scorer in §3.5.1 and Appendix A. The evaluation therefore largely measures self-consistency among LLM outputs rather than retrieval quality. Moreover, Definition 2 counts any tool in the same leaf sub-category as a hit; with K=1 at the pruning levels, a query can be 'successful' even if the exact intended tool is not returned. This does not substantiate the claim of 'matching state-of-the-art retrieval accuracy.' Independent human relevance judgments and an exact-tool or properly ranked hit metric are needed.
  3. [§3.2.1 Remark 2, §5.3] The O(log N) claim is not established. Remark 2 states the queried-tool count as O(βh + S), where S is the size of the leaf node's tool list; S is not bounded by any logarithmic quantity in the design. The two-level experiment in §5.3 reports an average leaf candidate set of about 1,197 tools at N ≈ 25,300, which is not asymptotically logarithmic. No balanced-tree assumption, branching-factor distribution, or depth measurements are provided. The 95.26% reduction is an empirical constant-factor improvement at one scale, not evidence of O(log N) scaling. The asymptotic claim needs either a formal bound on S and β or a revised statement.
minor comments (4)
  1. [Table 1] The final row is garbled: 'Remaining' is shown as '100' and the Step 6 rate is '18.14%'. The raw counts (54,700 → 33,688 after removing 21,012) should be presented consistently, and the column headers should be aligned with the steps.
  2. [§3.1] 'As illustrated in Figure ?' contains a missing figure reference; the figure number should be filled in.
  3. [Appendix A] The prompt templates contain garbled text: 'öther¨' should be 'other', and several curly braces in the prompt placeholders are inconsistently formatted.
  4. [General] Typographical issues include 'offical', 'ereinafter', and inconsistent 'REST-ful' / 'RESTful'. Please run a proofreading pass.

Circularity Check

2 steps flagged

The reported retrieval accuracy is a self-consistency test: Qwen builds the taxonomy, labels, and ground-truth queries while ToolDNS's LLM scorer is judged by agreement with Qwen's own categories.

specific steps
  1. fitted input called prediction [Section 5.1, Step 7; Section 5.2; Appendix A]
    "Following the methodology of ToolLLM, we used Qwen to generate task queries for each tool. These “intent-tool” pairs serve as the ground truth to test the retrieval effectiveness of ToolDNS. ... The System prompt used in this experiment is provided as “Agent Role of TLD server” and “Agent Role of NS server” in Appendix.A."

    ToolDNS's semantic pruning is performed by an LLM ranker (Section 3.5.1: 'we use small LLM scoring'; Appendix A gives the LLM ranking prompts). The test queries and their ground-truth tool labels are generated by Qwen in the same pipeline that builds the dataset. The reported hit rate therefore measures the ranker's agreement with Qwen's own query-tool assignments, not retrieval accuracy against an independent or human-labeled ground truth. This is a fitted input being presented as a prediction: the evaluator and the evaluated share the label generator.

  2. self definitional [Section 5.1, Step 6; Section 5.2 (Definition 2)]
    "We again used Qwen to assign sub-categories to data within each major category and removed unclassifiable items. ... A retrieval is counted as a hit if the returned tool belongs to the same semantic sub-category as T."

    Definition 2 relativizes success to R_t, the set of tools in the same leaf sub-category. Those sub-categories were assigned by Qwen in Step 6, while ToolDNS's hierarchical pruning returns a sub-category selected by an LLM at each level. The hit-rate metric is thus defined in terms of the same classifier output that the experiment claims to validate; a 'hit' means the LLM scorer chose the LLM-assigned label. The accuracy comparison with ToolLLM is therefore a comparison of two models' agreement with Qwen's taxonomy, not an outcome of the DNS mechanism.

full rationale

The circularity is concentrated in the empirical validation, not in the DNS-based architecture itself. ToolDNS's O(log N) analysis (Remark 2) is a standard tree-search bound and is not circular. The 95.26% search-space number is an arithmetic consequence of the authors' own taxonomy's leaf sizes; while it is not independently validated, it is not itself a hidden reuse of the result. What makes the headline 'matching state-of-the-art retrieval accuracy' circular is the evaluation loop: Qwen generated the summaries, sub-category labels, and the query-tool ground truth (Section 5.1, Steps 1, 6, 7), and the ToolDNS scorer is an LLM prompt (Appendix A) judged by whether it lands in the Qwen-assigned sub-category (Section 5.2). The hit-rate metric is therefore a measure of an LLM's agreement with itself, so the central accuracy claim reduces to a self-consistency check. This warrants 7 rather than 8 because the framework has independent architectural content and the self-consistency could in principle have failed. The unmodified-resolver contradiction (Algorithm 1's cursor replacement vs. Section 3.3.2's compatibility claim) is a serious correctness issue but not a circularity, so it is not scored here. No load-bearing self-citation was found.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 3 invented entities

The central design relies on three ad hoc additions: a hypothetical .tools TLD, a new EDNS0 option, and a special label-cursor convention. The empirical results depend on several hand-chosen parameters (K, taxonomy depth, LLM-generated summaries). The paper does not provide independent evidence for any of these inventions outside the simulation setup.

free parameters (3)
  • K (top-K at each pruning step) = 1
    The evaluation sets K=1 for the highest pruning ratio (Section 5.3); this choice directly determines the reported 95.26% reduction and is not varied systematically.
  • Taxonomy depth and structure = 2 levels, 14 top-level categories (Appendix B)
    The benchmark's two-level hierarchy is created by the authors via LLM and manual labor; the measured search-space reduction is a direct artifact of this particular taxonomy, not a parameter-free design.
  • Semantic summaries per subdomain = LLM-generated keywords/sentences
    Each node's summary is produced offline by an LLM and used as the matching basis; these summaries are not derived from first principles and determine the pruning outcome.
axioms (3)
  • domain assumption Tool functionality can be adequately captured by a hierarchical taxonomy (Section 4.3).
    The entire pruning mechanism assumes that all tools are classifiable into a tree of functional categories; the paper acknowledges this breaks for 'highly niche tools' and would degrade to a centralized solution.
  • domain assumption Standard DNS resolvers will forward EDNS0 options and accept the partial-name referral pattern described in Algorithm 1 (Sections 3.3.2, 3.3.3).
    This is load-bearing for the 'no resolver modification' claim. Standard DNS referral semantics require the QNAME to be a descendant of the returned NS zone, which is not the case for the partial-name queries, so unmodified resolvers would not follow them.
  • standard math Balanced tree and bounded branching factor for the O(log N) complexity claim (Remark 2).
    The formal complexity statement O(beta * h + S) reduces to O(log N) only when the taxonomy is balanced and beta is constant; the paper does not demonstrate this for real tool distributions.
invented entities (3)
  • Reserved .tools top-level domain no independent evidence
    purpose: Semantic root for tool discovery; requires delegation by IANA or a suitable authority.
    No real deployment or prototype is shown; it is a hypothetical TLD requiring external governance action.
  • New EDNS0 option code for intent payload no independent evidence
    purpose: Carry natural-language query intent inside DNS requests.
    The paper defines a binary layout (Figure 4) but does not register an option code or implement it in a real resolver; no interoperability evidence exists.
  • Partially unfolded domain name cursor (underscore placeholder) no independent evidence
    purpose: Encodes search progress in the DNS name itself.
    This is a nonstandard use of DNS labels that requires servers and resolvers to interpret the placeholder; no implementation demonstrates that unmodified resolvers can process it.

pith-pipeline@v1.3.0-alltime-deepseek · 17306 in / 11563 out tokens · 109040 ms · 2026-08-02T15:54:26.972427+00:00 · methodology

0 comments
read the original abstract

The coming era of autonomous AI agents demands a discovery mechanism capable of navigating millions of tools, yet existing solutions buckle under O(N) complexity and centralized governance. Instead of building another fragile overlay, we propose ToolDNS, a radical framework that retrofits semantic tool discovery onto the Internet's most resilient substrate: the Domain Name System (DNS). By embedding functional intent and organizational trust into a hierarchical namespace, ToolDNS transforms an expensive semantic search into a series of lightweight, O(log N) name resolutions. We introduce three protocol-compliant enhancements to enable decentralized governance and semantic pruning: partially unfolded names, EDNS0 intent payloads, and logical subdomains. To rigorously evaluate this approach across the fragmented tooling landscape, we construct and release a large-scale heterogeneous benchmark comprising 33,688 real-world tools spanning MCP, A2A, RESTful, and Skill protocols. On this dataset, ToolDNS slashes the per-query search space by 95.26% while matching state-of-the-art retrieval accuracy. Furthermore, its UDP-native design reduces discovery latency by orders of magnitude compared to HTTP-based registries. Our work demonstrates that scalable AI interoperability requires not more middleware, but a smarter utilization of the infrastructure already beneath our feet.

Figures

Figures reproduced from arXiv: 2607.18242 by Enhao Chen, Yulin Shao.

Figure 1
Figure 1. Figure 1: Existing discovery paradigms (left) versus the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Hierarchical domain space architecture of ToolDNS. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Logic subdomain structure: From the TLD entry [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Packet format (payload starts at bit 32). [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The comprehensive architecture and mechanism of [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Caching Mechanism of Recursive Resolver. This [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Retrieval accuracy comparison: ToolDNS versus [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Scalability of search space size. The comparison [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Retrieval accuracy of hierarchical versus flat orga [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 7 linked inside Pith

  1. [1]

    Roy Arends, Rob Austein, Matt Larson, Dan Massey, and Scott Rose. 2005. RFC 4033: DNS security introduction and requirements

    2005

  2. [2]

    2005.Reliable distributed systems: technologies, web services, and applications

    Kenneth P Birman. 2005.Reliable distributed systems: technologies, web services, and applications. Springer

    2005

  3. [3]

    Enfang Cui, Yujun Cheng, Rui She, Dan Liu, Zhiyuan Liang, Minxin Guo, Tianzheng Li, Qian Wei, Wenjuan Xing, and Zhijie Zhong. 2025. AgentDNS: A Root Domain Naming System for LLM Agents. arXiv:2505.22368 [cs.AI] https://arxiv.org/abs/2505.22368

    Pith/arXiv arXiv 2025

  4. [4]

    Hongwei Cui, Yuyang Du, Qun Yang, Yulin Shao, and Soung Chang Liew. 2024. LLMind: Orchestrating AI and IoT with LLM for complex task execution.IEEE Communications Magazine63, 4 (2024), 214–220

    2024

  5. [5]

    Joao Damas, Michael Graff, and Paul Vixie. 2013. RFC 6891: Extension mecha- nisms for DNS (EDNS (0))

    2013

  6. [6]

    Yunus Durmus and Ertan Onur. 2015. Service knowledge discovery in smart machine networks.Wireless Personal Communications81, 4 (2015), 1455–1480

    2015

  7. [7]

    Abul Ehtesham, Aditi Singh, Gaurav Kumar Gupta, and Saket Kumar. 2025. A survey of agent interoperability protocols: Model context protocol (MCP), agent communication protocol (ACP), agent-to-agent protocol (A2A), and agent network protocol (ANP).arXiv preprint arXiv:2505.02279(2025)

    Pith/arXiv arXiv 2025

  8. [8]

    Robert Elz and Randy Bush. 1997. RFC2181: Clarifications to the DNS Specifica- tion

    1997

  9. [9]

    2000.Architectural styles and the design of network-based software architectures

    Roy Thomas Fielding. 2000.Architectural styles and the design of network-based software architectures. University of California, Irvine

    2000

  10. [10]

    De Filippi et al. 2016. The invisible politics of Bitcoin: governance crisis of a decentralised infrastructure.Internet Policy Review5, 3 (2016)

    2016

  11. [11]

    Arnt Gulbrandsen, Paul Vixie, and Levon Esibov. 2000. RFC2782: A DNS RR for specifying the location of services (DNS SRV)

    2000

  12. [12]

    Mark Handley. 2006. Why the Internet only just works.BT Technology Journal 24, 3 (2006), 119–129

    2006

  13. [13]

    Paul Hoffman and Patrick McManus. 2018. RFC 8484: DNS queries over HTTPS (DoH)

    2018

  14. [14]

    Zi Hu, Liang Zhu, John Heidemann, Allison Mankin, Duane Wessels, and Paul Hoffman. 2016. RFC 7858: Specification for DNS over transport layer security (TLS)

    2016

  15. [15]

    Ken Huang, Vineeth Sai Narajala, Idan Habler, and Akram Sheriff. 2026. Agent name service (ANS): A universal directory for secure AI agent discovery and interoperability. InInternational Conference on AI in Cybersecurity (ICAIC). IEEE, 1–9

    2026

  16. [16]

    Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. 2024. Go- rilla: Large language model connected with massive APIs.Advances in Neural Information Processing Systems37 (2024), 126544–126565

    2024

  17. [17]

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. 2023. ToolLLM: Facilitating large language models to master 16000+ real-world APIs.arXiv preprint arXiv:2307.16789(2023)

    Pith/arXiv arXiv 2023

  18. [18]

    Ramesh Raskar, Pradyumna Chari, John Zinky, Mahesh Lambe, Jared James Grogan, Sichao Wang, Rajesh Ranjan, Rekha Singhal, Shailja Gupta, Robert Lincourt, et al. 2025. Beyond DNS: Unlocking the internet of AI agents via the nanda index and verified agentfacts.arXiv preprint arXiv:2507.14263(2025)

    Pith/arXiv arXiv 2025

  19. [19]

    Partha Pratim Ray. 2025. A survey on model context protocol: Architecture, state-of-the-art, challenges and future directions.Authorea Preprints(2025)

    2025

  20. [20]

    David P Reed. 2010. End-to-end arguments: The Internet and beyond. InUSENIX Security Symposium

    2010

  21. [21]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools.Advances in Neural Information Processing Systems36 (2023), 68539–68551

    2023

  22. [22]

    Yulin Shao, Qi Cao, and Deniz Gündüz. 2024. A theory of semantic communica- tion.IEEE Transactions on Mobile Computing23, 12 (2024), 12211–12228

    2024

  23. [23]

    Yulin Shao, Deniz Gündüz, and Soung Chang Liew. 2021. Federated edge learn- ing with misaligned over-the-air computation.IEEE Transactions on Wireless Communications21, 6 (2021), 3951–3964

    2021

  24. [24]

    Ion Stoica and Scott Shenker. 2021. From cloud computing to sky computing. In Proceedings of the Workshop on Hot Topics in Operating Systems. 26–32

    2021

  25. [25]

    OpenClaw Team. 2024. OpenClaw Documentation. https://docs.openclaw.ai/. Accessed: 2026-03-30

    2024

  26. [26]

    Qwen Team. 2025. Qwen3 Technical Report. arXiv:2505.09388 [cs.CL] https: //arxiv.org/abs/2505.09388

    Pith/arXiv arXiv 2025

  27. [27]

    Lukas Weidener, Marko Brkić, Phillip Lee, Martin Karlsson, Kevin Noessler, and Paul Kohlhaas. 2026. From agent-only social networks to autonomous scientific research: Lessons from OpenClaw and Moltbook, and the architecture of ClawdLab and Beach.Science.arXiv preprint arXiv:2602.19810(2026)

    arXiv 2026

  28. [28]

    Niklaus Wirth. 2002. A plea for lean software.Computer28, 2 (2002), 64–68

    2002

  29. [29]

    Mengying Wu, Pei Chen, Geng Hong, Baichao An, Jinsong Chen, Binwang Wan, Xudong Pan, Jiarun Dai, and Min Yang. 2025. MCPZoo: A Large-Scale Dataset of Runnable Model Context Protocol Servers for AI Agent.arXiv preprint arXiv:2512.15144(2025)

    arXiv 2025

  30. [30]

    Renjun Xu and Yang Yan. 2026. Agent skills for large language models: Architec- ture, acquisition, security, and the path forward.arXiv preprint arXiv:2602.12430 (2026)

    Pith/arXiv arXiv 2026

  31. [31]

    John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024. Swe-agent: Agent-computer interfaces enable automated software engineering.Advances in Neural Information Processing Systems37 (2024), 50528–50652

    2024

  32. [32]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations

    2022

  33. [33]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models.arXiv preprint arXiv:2303.182231, 2 (2023), 1–124. 13

    Pith/arXiv arXiv 2023