Pith. sign in

REVIEW 3 major objections 5 minor 47 references

HAWK: A Hierarchical Workflow Framework for Multi-Agent Collaboration

T0 review · 3 major / 5 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 A five-layer workflow framework aims to make multi-agent AI systems cooperate through standardized interfaces.

desk verdict A plausible layered framework whose central adaptive-scheduling claims aren't yet tested: the optimizers are unimplemented, and the CreAgentive experiment has no baselines. read the letter →

arxiv 2507.04067 v1 pith:CFPNFA67 submitted 2025-07-05 cs.AI cs.MA

Yuyang Cheng , Yumiao Xu , Chaojia Yu , Yong Zhao This is my paper · ORCID
classification cs.AIcs.MA
keywords multi-agentsystemsworkflowframeworkhierarchicalarchitectureadaptiveschedulingresourceabstractionstandardizedinterfaceslargelanguagemodelsnovelgeneration
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 proposes HAWK, a five-layer framework for building multi-agent AI systems, with User, Workflow, Operator, Agent, and Resource layers connected by sixteen standardized interfaces. The central claim is that this layered separation makes heterogeneous agents interoperable, replaces static scheduling with adaptive scheduling, and unifies access to data, models, devices, and tools. As evidence, the authors built CreAgentive, a multi-agent novel-writing prototype, and report that it runs a ten-chapter story in about eighty minutes, keeps over ninety-two percent of modules stable across fifty runs, and supports five parallel storylines. The paper acknowledges that the Workflow Optimizer and Task Optimizer modules that would deliver the adaptive scheduling are not yet implemented, so the unique advantage rests on components still to be built.

What carries the argument

The load-bearing object is the reference architecture itself: five layers (User, Workflow, Operator, Agent, Resource), twenty modules, and sixteen interfaces (I1–I16) that standardize communication between the modules. The mechanism that carries the argument is the adaptive scheduling and optimization loop: the Workflow Engine receives real-time execution states via Interface I4, the Workflow Optimizer uses that feedback to adjust workflow structure and scheduling through Interface I2, and the Task Optimizer in the Operator Layer adjusts strategies mid-execution. In CreAgentive, the Decision Agent adds a differentiable Disjunctive Normal Form layer over ten narrative predicates to rank candidate storylines, which is the one component that is fully implemented and tested.

What would settle it

Run HAWK on a workload whose task types and resource constraints change mid-execution, and compare throughput and resource utilization against a static scheduling baseline; if the optimizer modules are absent or give no measurable gain, the adaptive-scheduling claim is unsupported.

Watch

Extended reading notes

Core claim

HAWK's central claim is architectural: organizing an agent system into five decoupled layers with sixteen defined interfaces lets each layer evolve independently while preserving end-to-end control from user request to resource invocation. In the Workflow Layer, real-time feedback from workflow monitoring is supposed to feed a Workflow Optimizer that adjusts workflow structure and scheduling strategy dynamically; in the Operator Layer, a Task Optimizer adjusts execution strategy based on policies and available resources. The Resource Layer abstracts heterogeneous data, large models, physical devices, and third-party tools behind one access mechanism. The authors claim that applying this design in CreAgentive improves task throughput, simplifies resource invocation, and enhances system controllability, and that hybrid deployments of large language models integrate cleanly.

Load-bearing premise

The framework's promised advantage over existing systems rests on the Workflow Optimizer and Task Optimizer modules, which are described but not implemented in the prototype; if those modules cannot perform real-time dynamic adjustment, HAWK becomes a standard layered architecture.

Editorial extensions

If this is right

  • If HAWK's layered interfaces are adopted, new agents and tools can be added without rewriting scheduling or resource code, since they plug in at the Agent and Resource layers.
  • If the adaptive optimizer works as specified, systems can shift tasks to available resources in real time instead of relying on hand-coded static rules.
  • The unified Resource Layer means an application can call a local database, a cloud large language model, a robotic arm, or a third-party API through the same access pattern.
  • The CreAgentive results imply that the same workflow shell can be reused for other creative generation tasks, such as screenwriting or interactive narrative.
  • The hybrid-LLM comparison suggests that routing different workflow stages to different models can beat using one model for everything.

Reading between the lines

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

  • A controlled benchmark that varies task types and resource constraints mid-execution, with and without the optimizer modules, would test whether the adaptive-scheduling claim holds; the current prototype does not report such a comparison.
  • If the optimizer modules fail to deliver, HAWK's differentiation reduces to a conventional layered architecture whose value lies mainly in interface standardization rather than intelligence.
  • The positioning over MCP, ANP, and A2A suggests HAWK could serve as a compatibility layer on top of those protocols, but the paper offers no implementation demonstrating that convergence.
  • The same interface set might extend to non-LLM agent domains such as robotic fleets, since physical devices are already a first-class Resource category; a testable extension would be a hardware-in-the-loop deployment.
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.

Referee Report

3 major / 5 minor

Summary. This manuscript proposes HAWK, a five-layer reference architecture (User, Workflow, Operator, Agent, Resource) with sixteen standardized interfaces for multi-agent collaboration, and describes CreAgentive, a novel-generation prototype built on a subset of HAWK. The paper's central contributions are the architecture itself, its claimed adaptive scheduling and optimization modules, and an empirical demonstration of throughput and controllability improvements. It also includes a capability comparison of existing frameworks and a discussion of application domains, future directions, and failure modes.

Significance. If fully implemented and validated, HAWK would provide a useful organizing reference model for multi-agent systems: the layered decomposition and interface enumeration are clear, and the paper honestly marks unimplemented modules and known failure modes in Section 4.3. However, as submitted, the core adaptive scheduling and optimization mechanism is present only as a design, and the only evaluation is a descriptive case study with no baselines, error bars, or comparisons to existing frameworks. The contribution is best assessed as an early architecture proposal rather than a validated framework, and the empirical claims in the abstract and conclusion outrun the evidence presented.

major comments (3)
  1. [§4.3 and §5] The central claim that HAWK provides adaptive scheduling and optimization is not supported by the implementation. Figure 1 marks the Workflow Monitoring, Workflow Optimizer, and Task Optimizer as gray (to be developed), and Section 4.3 states that 'components shaded in gray remain to be developed'; consequently, CreAgentive does not exercise Interface I2, the Task Optimizer, or the Workflow Optimizer. The Section 5 sentence 'HAWK significantly improves task throughput, simplifies resource invocation, and enhances overall system controllability' therefore attributes to HAWK a mechanism that was never run in the reported experiment.
  2. [§4.3] The experimental support consists of descriptive statistics without baselines or error bars: 80 minutes per 10-chapter story, 92% module stability across 50 runs, and 5 concurrent storylines. No comparison is made to a static scheduler, to a non-HAWK pipeline, or to any existing multi-agent framework, and no variance or significance testing is reported; these figures cannot establish 'significant' improvements in throughput, invocation complexity, or controllability.
  3. [§2, Table 1] The capability comparison in Table 1 is derived from an unpublished manuscript (reference [38]), and no methodology or per-cell evidence is given for the ✓/✗/# ratings. Without reproducible sources or a defined rating procedure, the heatmap cannot be independently checked, and it is therefore weak evidence for the claimed advantages of HAWK over existing systems.
minor comments (5)
  1. [§3.1 and §3.3] There are typographical errors: 'Wrkflow Engine' appears in Section 3.1, and 'modular abstraction and and supports' appears in Section 3.3 with a duplicated 'and'.
  2. [§2, Tables 1–2] The symbol legend in Table 2 does not explain the combined symbols that appear in Table 1 cells (for example, the cells rendered as 'H #' in AutoGen, Coze, DeepResearch, Flowise, and other rows), and the table does not distinguish open-source frameworks, commercial products, and research prototypes.
  3. [§4.3] The cross-model comparison (DeepSeek-V3, Qwen/QwQ-32B, GLM-4-9B) is reported without controlling for prompt format, token budget, decoding parameters, or hardware; the resulting 'Average Chapters' and qualitative weakness labels are anecdotal rather than controlled measurements.
  4. [§4 and §5] No code, repository, or configuration details are provided for CreAgentive, which limits reproducibility of the 50-run stability evaluation and the 80-minute runtime figure.
  5. [§3.3] The assertion that HAWK 'fully encompasses' the capabilities of MCP, ANP, and A2A is made without a concrete mapping between the sixteen HAWK interfaces and the corresponding protocol operations, so the claim is not checkable.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: HAWK's central claim is an empirical demonstration, not a derivation from fitted inputs; the only overlapping-author citation is not load-bearing.

full rationale

The paper does not derive any quantitative prediction from a fitted parameter, and it does not define its claimed outputs in terms of its inputs. The central claim in Section 5 — that HAWK 'significantly improves task throughput, simplifies resource invocation, and enhances overall system controllability' — rests on the CreAgentive prototype described in Section 4, and the reported metrics (80 minutes per 10-chapter story, 92% module stability, 5 concurrent storylines) are descriptive observations from that prototype. These may be weak evidence with no baselines or error bars, and Section 4.3 explicitly states that components 'shaded in gray remain to be developed,' meaning the Workflow Optimizer and Task Optimizer are not exercised in the experiment. That is a correctness/evidence gap, not circularity: the claim does not reduce to the framework's definitions by construction. The only overlapping-author citation is reference [38], used for the capability comparison in Table 1 ('Drawing on the work of Yu et al. [38], We have summarized and compared the functionalities of existing multi-agent frameworks in Table 1'). This citation is self-referential in authorship, but it is not load-bearing for the paper's central contribution: the comparison table is contextual positioning, and HAWK's architecture and CreAgentive demonstration stand independently of that survey. There is no imported uniqueness theorem, no ansatz smuggled via citation, and no renaming of a known result as unification. The DNF reasoning equations in Section 4.2 are implementation details, not a derivation of HAWK's effectiveness. Therefore no circular step can be exhibited, and the appropriate finding is no significant circularity.

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

The paper rests on the design assumption that layered abstraction and interface standardization yield real benefits, and on the practical assumption that LLMs and the adapted DNF module work well enough for the prototype. Neither is given strong evidence.

assumptions (4)
  • domain assumption A layered architecture with standardized interfaces improves interoperability, scalability, and controllability of multi-agent systems.
    The central design premise of HAWK; the paper states it in Section 3 but provides no empirical or formal validation.
  • domain assumption Recent LLMs can reliably execute the agent roles in the prototype, including goal generation, planning, writing, and ending checking, at sufficient quality.
    The prototype's function depends on LLM capabilities; the paper only reports limited anecdotal comparisons in Section 4.3.
  • domain assumption The DNF reasoning layer from TELLER can be effectively adapted to evaluate and rank candidate storylines.
    Assumed without independent validation in this context; see Section 4.2.
  • domain assumption The capability heatmap in Table 1, based on the unpublished survey by the same research group, accurately represents the features of each listed system.
    Table 1 underpins the claimed gap in the ecosystem; it relies on an unpublished manuscript (reference [38]).

how reviews work

0 comments
Cite this review

Pith. "Pith review of HAWK: A Hierarchical Workflow Framework for Multi-Agent Collaboration." pith.science (2026). https://pith.science/paper/CFPNFA67

@misc{pith2026250704067,
  author       = {Pith},
  title        = {Pith review of: HAWK: A Hierarchical Workflow Framework for Multi-Agent Collaboration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CFPNFA67}},
  note         = {Machine review of arXiv:2507.04067}
}
read the original abstract

Contemporary multi-agent systems encounter persistent challenges in cross-platform interoperability, dynamic task scheduling, and efficient resource sharing. Agents with heterogeneous implementations often lack standardized interfaces; collaboration frameworks remain brittle and hard to extend; scheduling policies are static; and inter-agent state synchronization is insufficient. We propose Hierarchical Agent Workflow (HAWK), a modular framework comprising five layers-User, Workflow, Operator, Agent, and Resource-and supported by sixteen standardized interfaces. HAWK delivers an end-to-end pipeline covering task parsing, workflow orchestration, intelligent scheduling, resource invocation, and data synchronization. At its core lies an adaptive scheduling and optimization module in the Workflow Layer, which harnesses real-time feedback and dynamic strategy adjustment to maximize utilization. The Resource Layer provides a unified abstraction over heterogeneous data sources, large models, physical devices, and third-party services&tools, simplifying cross-domain information retrieval. We demonstrate HAWK's scalability and effectiveness via CreAgentive, a multi-agent novel-generation prototype, which achieves marked gains in throughput, lowers invocation complexity, and improves system controllability. We also show how hybrid deployments of large language models integrate seamlessly within HAWK, highlighting its flexibility. Finally, we outline future research avenues-hallucination mitigation, real-time performance tuning, and enhanced cross-domain adaptability-and survey prospective applications in healthcare, government, finance, and education.

Figures

Figures reproduced from arXiv: 2507.04067 by the authors.

Figure 1
Figure 1. Framework of Hierarchical Agent WorKflow (HAWK) [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The Workflow of the CreAgentive In this section, we demonstrate the applicability of HAWK by implementing a multi-agent, AI-driven text generation prototype named CreAgentive. Drawing inspiration from established systems such as Drama Engine [42] and Multiagent-poetry [43], CreAgentive showcases how HAWK can effectively orchestrate a collection of specialized agents to collaboratively compose a novel. The workflow o… view at source ↗
Figure 3
Figure 3. illustrates the overall decision-making pipeline: logical truth values µi,k are evaluated through learned conjunctive and disjunctive structures, enabling interpretable, controllable selection of the optimal plot trajectory. • Writer Agent: The Writer Agent is tasked with transforming abstract plans into coherent narrative text. It takes as input the optimal trajectory selected by the Decision Module along with the … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 38 canonical work pages

  1. [38]

    Multi-agent deep reinforcement learning for urban traffic light control in vehicular networks

    Tong Wu, Pan Zhou, Kai Liu, Yali Yuan, Xiumin Wang, Huawei Huang, and Dapeng Oliver Wu. Multi-agent deep reinforcement learning for urban traffic light control in vehicular networks. IEEE Transactions on Vehicular Technology, 69(8):8243–8256, 2020

    work page 2020

  2. [1]

    Ai agents under threat: A survey of key security challenges and future pathways

    Zehang Deng, Yongjian Guo, Changzhou Han, Wanlun Ma, Junwu Xiong, Sheng Wen, and Yang Xiang. Ai agents under threat: A survey of key security challenges and future pathways. ACM Computing Surveys, 57(7):1–36, 2025

    work page 2025

  3. [2]

    agentuniverse: Multi-agent ecosystem for autonomous agents

    agentUniverse Contributors. agentuniverse: Multi-agent ecosystem for autonomous agents. https: //github.com/agentuniverse-ai/agentUniverse, 2024. Accessed: 2025-04-14

    work page 2024

  4. [3]

    Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents

    Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi Lu, Yi-Hsin Hung, Chen Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents. arXiv preprint arXiv:2308.10848, 2023. Accessed: 2025-04-20

    arXiv 2023

  5. [4]

    Agno: An agentic workflow framework for llm applications, 2024

    Agno Contributors. Agno: An agentic workflow framework for llm applications, 2024. Accessed: 2025-04-14

    work page 2024

  6. [5]

    Autogen: Enabling next-gen llm applications via multi-agent conversation framework

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. Accessed: 2025-04-20

    arXiv 2023

  7. [6]

    Camel: Communicative agents for "mind" exploration of large language model society

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for "mind" exploration of large language model society. arXiv preprint arXiv:2303.17760, 2023. Accessed: 2025-04-20

    arXiv 2023

  8. [7]

    Chatdev: Communicative agents for software development

    Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. Chatdev: Communicative agents for software development. arXiv preprint arXiv:2307.07924, 2023. Accessed: 2025-04-20

    arXiv 2023

Show all 47 references
  1. [8]

    Coze: Next-gen ai chatbot & multi-agent workflow platform

    ByteDance. Coze: Next-gen ai chatbot & multi-agent workflow platform. https://www.coze.com/,

  2. [9]

    Crewai: The leading multi-agent platform

    CrewAI Inc. Crewai: The leading multi-agent platform. https://github.com/crewAIInc/crewAI,

  3. [10]

    Introducing deep research: Openai’s new system for collaborative reasoning

    OpenAI. Introducing deep research: Openai’s new system for collaborative reasoning. OpenAI Blog,

  4. [12]

    Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts

    Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vard- hamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. Dspy: Compiling declarative language model calls into self-imp...

    arXiv

  5. [13]

    Ernie bot agent: Llm agent framework for baidu ernie

    PaddlePaddle. Ernie bot agent: Llm agent framework for baidu ernie. https://github.com/ PaddlePaddle/ERNIE-SDK, 2023. Accessed: 2025-04-20

    2023

  6. [14]

    Dify: Open-source llm application development platform

    langgenius. Dify: Open-source llm application development platform. https://github.com/ langgenius/dify, 2025. Accessed: 2025-04-20. 13

    2025

  7. [15]

    Exploration of llm multi-agent application implementation based on langgraph+crewai

    Zhihua Duan and Jialin Wang. Exploration of llm multi-agent application implementation based on langgraph+crewai. arXiv preprint arXiv:2411.18241, 2024. Accessed: 2025-04-20

    2024 arXiv

  8. [16]

    Magentic- one: A generalist multi-agent system for solving complex tasks

    Adam Fourney, Gagan Bansal, Hussein Mozannar, Cheng Tan, Eduardo Salinas, Erkang Zhu, Friederike Niedtner, Grace Proebsting, Griffin Bassman, Jack Gerrits, Jacob Alber, Peter Chang, Ricky Loynd, Robert West, Victor Dibia, Ahmed Awadallah, Ece Kamar, Rafah Hosn, and Saleema Ame...

    2024 arXiv

  9. [17]

    Flowise: Drag & drop ui to build llm apps with langchain

    FlowiseAI Contributors. Flowise: Drag & drop ui to build llm apps with langchain. https://github. com/FlowiseAI/Flowise, 2024. Accessed: 2025-04-14

    2024

  10. [18]

    n8n: Fair-code workflow automation platform with native ai capabilities

    n8n.io. n8n: Fair-code workflow automation platform with native ai capabilities. https://n8n.io,

  11. [19]

    Omagent: A multi-modal agent framework for complex video understanding with task divide-and-conquer

    Lu Zhang, Tiancheng Zhao, Heting Ying, Yibo Ma, and Kyusong Lee. Omagent: A multi-modal agent framework for complex video understanding with task divide-and-conquer. https://arxiv.org/ abs/2406.16620, 2024. Accessed: 2025-04-20

    2024 arXiv

  12. [20]

    Metagpt: Meta programming for a multi-agent collaborative framework

    Sirui Hong, Mingchen Zhuge, Jiaqi Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. Metagpt: Meta programming for a multi-agent collaborative frame...

    2023 arXiv

  13. [21]

    Manus: A general ai agent bridging thoughts and actions

    Monica (Butterfly Effect AI). Manus: A general ai agent bridging thoughts and actions. https: //manus.im/, 2025. Accessed: 2025-04-23

    2025

  14. [22]

    Accessed: 2025-04-20

    2025

  15. [23]

    Semantic kernel: Lightweight, open-source ai agent sdk

    Microsoft. Semantic kernel: Lightweight, open-source ai agent sdk. https://github.com/ microsoft/semantic-kernel, 2024. Accessed: 2025-04-20

    2024

  16. [24]

    Swarm: Educational framework exploring ergonomic, lightweight multi-agent orchestration

    OpenAI. Swarm: Educational framework exploring ergonomic, lightweight multi-agent orchestration. https://github.com/openai/swarm, 2024. Accessed: 2025-04-20

    2024

  17. [25]

    Big data workflows: A reference architecture and the dataview system

    Andrey Kashlev, Shiyong Lu, and Aravind Mohan. Big data workflows: A reference architecture and the dataview system. Services Transactions on Big Data (STBD), 4(1):1–19, 2017

    2017

  18. [26]

    Qwen-agent: Agent framework and applications for qwen models

    QwenLM. Qwen-agent: Agent framework and applications for qwen models. https://github.com/ QwenLM/Qwen-Agent, 2025. Accessed: 2025-04-20

    2025

  19. [27]

    Scientific workflow management and the kepler system

    Bertram Ludäscher, Ilkay Altintas, Chad Berkley, Dan Higgins, Efrat Jaeger, Matthew Jones, Edward A Lee, Jing Tao, and Yang Zhao. Scientific workflow management and the kepler system. Concurrency and computation: Practice and experience, 18(10):1039–1065, 2006

    2006

  20. [28]

    Collaborative scientific workflows

    Shiyong Lu and Jia Zhang. Collaborative scientific workflows. In 2009 IEEE International Conference on Web Services, pages 527–534. IEEE, 2009

    2009

  21. [29]

    A generic effi- cient scientific workflow engine for the optimizations of run-time execution

    Changxin Bai, Junwen Liu, Anik Tahabilder, MM Imran, Shiyong Lu, and Dunren Che. A generic effi- cient scientific workflow engine for the optimizations of run-time execution. In2023 IEEE International Conference on Software Services Engineering (SSE), pages 98–103. IEEE, 2023

    2023

  22. [30]

    Swift: Fast, reliable, loosely coupled parallel computation

    Yong Zhao, Mihael Hategan, Ben Clifford, Ian Foster, Gregor V on Laszewski, Veronika Nefedova, Ioan Raicu, Tiberiu Stef-Praun, and Michael Wilde. Swift: Fast, reliable, loosely coupled parallel computation. In 2007 IEEE Congress on Services (Services 2007), pages 199–206. IEEE, 2007

    2007

  23. [31]

    Tian and Y .D

    W.D. Tian and Y .D. Zhao. Optimized Cloud Resource Management and Scheduling: Theories and Practices. Morgan Kaufmann, 2014

    2014

  24. [32]

    Lpod: A local path based optimized scheduling algorithm for deadline-constrained big data workflows in the cloud

    Changxin Bai, Shiyong Lu, Ishtiaq Ahmed, Dunren Che, and Aravind Mohan. Lpod: A local path based optimized scheduling algorithm for deadline-constrained big data workflows in the cloud. In 2019 IEEE International Congress on Big Data (BigDataCongress), pages 35–44. IEEE, 2019. 14

    2019

  25. [33]

    An adaptive som neural network method for distributed formation control of a group of auvs

    Xin Li and Daqi Zhu. An adaptive som neural network method for distributed formation control of a group of auvs. IEEE Transactions on Industrial Electronics, 65(10):8260–8270, 2018

    2018

  26. [34]

    A service framework for scientific workflow management in the cloud

    Yong Zhao, Youfu Li, Ioan Raicu, Shiyong Lu, Cui Lin, Yanzhe Zhang, Wenhong Tian, and Ruini Xue. A service framework for scientific workflow management in the cloud. IEEE Transactions on Services Computing, 8(6):930–944, 2014

    2014

  27. [35]

    A survey on intelligent control for multiagent systems

    Peng Shi and Bing Yan. A survey on intelligent control for multiagent systems. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 51(1):161–175, 2020

    2020

  28. [36]

    Enhancing diagnostic capability with multi-agents conversational large language models

    Xi Chen, Huahui Yi, Mingke You, WeiZhi Liu, Li Wang, Hairui Li, Xue Zhang, Yingman Guo, Lei Fan, Gang Chen, et al. Enhancing diagnostic capability with multi-agents conversational large language models. NPJ digital medicine, 8(1):159, 2025

    2025

  29. [37]

    Huawei mindspore ai development framework

    Huawei Technologies Co., Ltd. Huawei mindspore ai development framework. In Artificial Intelligence Technology, pages 137–162. Springer Nature Singapore, 2023

    2023

  30. [39]

    Introducing the model context protocol

    Anthropic. Introducing the model context protocol. https://www.anthropic.com/news/ model-context-protocol, 2024. Accessed: 2025-04-23

    2024

  31. [40]

    Phidata: Framework for building multi-modal agents and workflows

    Phidata (Agno). Phidata: Framework for building multi-modal agents and workflows. https://docs. phidata.com/introduction, 2025. Accessed: 2025-04-20

    2025

  32. [41]

    Announcing the agent2agent protocol (a2a)

    Google. Announcing the agent2agent protocol (a2a). https://developers.googleblog.com/en/ a2a-a-new-era-of-agent-interoperability/ , 2025. Accessed: 2025-04-23

    2025

  33. [42]

    A survey on agent workflow - status and future

    Chaojia Yu, Zihan Cheng, Hanwen Cui, Yishuo Gao, Zexu Luo, Yijin Wang, Hangbin Zheng, and Yong Zhao. A survey on agent workflow - status and future. Unpublished manuscript, 2025. Available from the authors or at https://scupi.scu.edu.cn/faculty-staff/faculty/%e8%b5%b5%e5%8b%87

    2025

  34. [43]

    Llm-based multi-agent poetry generation in non-cooperative environments, 2024

    Ran Zhang and Steffen Eger. Llm-based multi-agent poetry generation in non-cooperative environments, 2024

    2024

  35. [44]

    Agent network protocol (anp): Building the http of the agentic web era

    Agent Network Protocol Community. Agent network protocol (anp): Building the http of the agentic web era. https://agentnetworkprotocol.com/en/, 2025. Accessed: 2025-04-23

    2025

  36. [45]

    Pan, Shuyi Yang, Lakshya A

    Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Dan Klein, Kannan Ramchandran, Matei Zaharia, Joseph E. Gonzalez, and Ion Stoica. Why do multi-agent llm systems fail?, 2025. 15

    2025

  37. [46]

    Drama engine: A framework for narrative agents, 2024

    Martin Pichlmair, Riddhi Raj, and Charlene Putney. Drama engine: A framework for narrative agents, 2024

    2024

  38. [48]

    TELLER: A trustworthy framework for explainable, generalizable and controllable fake news detection

    Hui Liu, Wenya Wang, Haoru Li, and Haoliang Li. TELLER: A trustworthy framework for explainable, generalizable and controllable fake news detection. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Findings of the Association for Computational Linguistics: ACL 2024, ...

    2024

  39. [2025]

    Accessed: 2025-04-14

    2025

Pith tools

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