Pith. sign in

REVIEW 4 major objections 3 minor 1 cited by

COALESCE: Economic and Security Dynamics of Skill-Based Task Outsourcing Among Team of Autonomous LLM Agents

T0 review · 4 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read LLM agents can save 20.3% on real tasks by outsourcing subtasks to specialized contractor agents—but only when the decision rule explores the market.

desk verdict Coherent framework, but the real-world cost savings are arithmetically impossible under the paper's own prices, and the simulation numbers are generated from the authors' own calibrated distributions. read the letter →

arxiv 2506.01900 v1 pith:EN5UQKC4 submitted 2025-06-02 cs.AI cs.CEcs.CR

classification cs.AIcs.CEcs.CR
keywords LLMagentstaskoutsourcingcostoptimizationmulti-agentsystemsepsilon-greedyexplorationagenteconomiesA2AprotocolTOPSISdecision-making
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

This paper claims that autonomous LLM agents can lower their operating costs by outsourcing resource-intensive subtasks to specialized contractor agents whenever the contractor's all-in price beats local execution. To make that decision, the paper introduces COALESCE, a framework that combines hybrid skill representation and discovery, task decomposition, a unified internal-versus-external cost model, a multi-criteria selection rule, and a standardized agent-to-agent communication flow. Validation across 239 simulated market runs reports a 41.8% average cost reduction, while 240 real LLM tasks show 20.3% cost reduction when the decision rule includes $\epsilon$-greedy exploration and only 1.9% when it does not. A sympathetic reading takes the core claim to be that a market for agent capabilities is economically viable and that exploration is the practical price of accessing it.

What carries the argument

The load-bearing object is the COALESCE decision engine: a unified cost model (equations 1–10) that prices local execution and contractor execution in the same units, a TOPSIS multi-criteria ranker with correlation-adjusted weights and a game-theoretic equilibrium check (Listing 1), and an $\epsilon$-greedy exploration gate ($\epsilon = 0.1$) that periodically selects a contractor outside the usual thresholds to discover market opportunities (Listing 2). A hybrid skill representation — ontology-style capability cards plus learned skill embeddings — feeds the compatibility filter, and the outsourcing workflow rides on the A2A agent-to-agent protocol's task lifecycle.

What would settle it

Take the 240-task validation and log every token sent to and received from each API, then recompute local execution cost as tokens times published per-token price plus fixed overhead; if the genuine local cost per task approaches the contractor prices, the 20.3% improvement will not reproduce.

Watch

Extended reading notes

Core claim

The central claim is that a client LLM agent can decide when to outsource by comparing modeled internal cost $C_{\text{internal}}(T)$ — compute, memory, energy, opportunity, depreciation — against external cost $C_{\text{external}}(T, A_j)$ — contractor price, communication, verification, integration, risk, latency — and that this comparison, run through skill-compatibility filtering and TOPSIS ranking, reliably selects outsourcing when it is economically beneficial. The paper reports 41.8% $\pm$ 10.5% cost reduction across 239 theoretical simulation runs, and in real API-based validation across 240 tasks reports that the same decision engine with $\epsilon = 0.1$ exploration achieves 20.3% cost reduction while disabling exploration leaves only 1.9%. The real-agent result is framed as showing both viability and a critical dependency: exploration is necessary to discover contractor relationships that the deterministic rule would never try.

Load-bearing premise

The load-bearing premise is that the 240 real tasks cost $0.00002 each to run locally and that $0.80–$2.00 contractor prices are the fair comparison; if that local baseline is wrong or omits real overhead, the reported 20.3% saving collapses.

Editorial extensions

If this is right

  • A deployment that skips exploration will quietly default to local execution; the paper's real-agent run shows savings collapsing from 20.3% to 1.9%, so exploration must be treated as a required component rather than a tuning option.
  • The unified cost model gives agents a common accounting language in which GPU-heavy subtasks such as large retrieval, fine-tuning, and batch inference can be priced and traded between agents.
  • Because simulated savings fall as the agent population grows (from 70.9% at 5 clients to 35.4% at 50), system designers should expect coordination overhead and market congestion to erode outsourcing gains at scale.
  • An inter-agent communication standard is sufficient substrate for an economic layer: the paper's workflow maps discovery, task initiation, result delivery, and verification onto the A2A lifecycle, with payment and reputation handled outside that protocol.

Reading between the lines

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

  • A useful extension is to transplant the same cost-comparison logic to open-weight local models, where itemized GPU-hour pricing would let a third party verify the margin without relying on proprietary API rates.
  • The large per-task savings reported during exploration suggest the main inefficiency is ignorance of contractor quality, so a market-maker or capability registry could capture much of the exploration benefit deterministically — a direction the paper itself lists among its mitigations.
  • One could also test whether a decaying exploration schedule, starting near 0.2 and falling to 0.02 as reputation data accumulates, preserves most of the 20.3% while reducing random contractor trials.
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

4 major / 3 minor

Summary. The paper proposes COALESCE, a framework for cost-aware task outsourcing among autonomous LLM agents. It combines hybrid skill representation, dynamic skill discovery, task decomposition, an internal/external cost model, TOPSIS-based selection with epsilon-greedy exploration, and integration with the A2A communication protocol. The central claims are a 41.8% ± 10.5% cost reduction across 239 theoretical simulations and a 20.3% cost reduction across 240 real LLM API tasks with epsilon-greedy exploration, versus only 1.9% without it. The manuscript also analyzes limitations and proposes mitigation strategies for exploration dependency.

Significance. If valid, COALESCE would address a real problem: many LLM deployments are GPU- and cost-constrained, and dynamic specialization could lower operating costs. The architectural proposals and the candid discussion of security, trust, and market-design issues are reasonable and potentially useful for future agent-market research. However, the quantitative validation does not survive scrutiny: the real-world number is arithmetically incompatible with the paper's own cost schedule, and the simulation number is generated from assumptions the authors selected. The paper is transparent about some of these limitations in Section VI, but the abstract and conclusion present the numbers as established results. No machine-checked proofs or reproducible experiment logs accompany the manuscript.

major comments (4)
  1. [Section V.I, Eq. (6)] The claimed 20.3% real-agent cost reduction is arithmetically inconsistent with the stated prices. The paper sets local execution at $0.00002/task and contractor prices at $0.80, $1.50, and $2.00/task. Since Eq. (6) adds communication, verification, integration, risk, and latency costs to the contractor price, every outsourced task costs at least $0.80 more than local execution. For 240 tasks, the local baseline is 240 × $0.00002 = $0.0048; outsourcing 11.4% of tasks (about 27 tasks) adds roughly 27 × $0.80 = $21.6, an increase of several orders of magnitude, not a 20.3% reduction. The 344.4% 'savings' on individual tasks is likewise impossible under the stated model. Section VI.G.1 itself concedes that the 'extreme cost differential' makes the algorithm rationally choose local execution. The paper must either redefine the cost metric, correct the local-cost baseline, or provide per-task cost breakdowns; as written, the central empirical claim is unsupported.
  2. [Section V.A and VI.A] The 41.8% theoretical validation cannot serve as validation of the framework. The simulation generates synthetic agent behavior through predefined probability distributions chosen by the authors, as VI.A explicitly states: 'The simulation framework generates synthetic agent behavior through predefined probability distributions rather than actual LLM agent interactions.' Costs, prices, and capabilities are calibrated from the authors' own distributions, so the simulation largely returns the assumptions put into it. The sensitivity analysis and aggregate statistics in Table I are therefore evidence about the simulator, not about real LLM agents.
  3. [Section V.E, Table I] The simulation results are internally inconsistent. The text states that '1-day simulations show limited performance (0% cost reduction)' and that '3-day and 20-day configurations achieve peak performance (55.4% and 98.0% cost reduction respectively),' but Table I reports dur_01 as 27.1% and contains no 20-day configuration. Additionally, 17 configurations × 20 runs per configuration = 340 runs, not the 239 runs aggregated in Table I; the paper does not explain which runs were excluded.
  4. [Section V.I, Table II] The real-agent validation numbers are internally inconsistent. The text reports 'Four confirmed HTTP requests' during exploration, while Table II reports 28 confirmed API calls; Table II also lists 35 tasks for the no-exploration run, while the text describes two validation runs over 240 tasks. The paper does not reconcile these counts or state how the 240-task total is composed across the two experiments.
minor comments (3)
  1. [Table I, Section V.B] The table header says '0.8 TOPSIS' but Section V.B sets the TOPSIS threshold τ_threshold = 0.6, and Section V.F.3 later says TOPSIS scores average 0.800; please clarify which value is intended.
  2. [Listing 1, line 24] The comment 'Apply correlation adjustment from Equation (1)' should reference Equation (11), which is the correlation-adjusted weight formula.
  3. [Abstract vs. Section I.C] The COALESCE acronym is expanded inconsistently as 'Cost-Optimized and Secure Agent Labour Exchange' in the abstract and 'Cost-Optimized Agent Labor Exchange' in Section I.C; please standardize.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline real-world validation is not a prediction from COALESCE's own cost model: Eq. (6) plus the stated $0.00002 local cost and $0.80–$2.00 contractor prices makes outsourcing strictly loss-making, and the 'exploration is essential' result is generated by the pre-calibrated ε.

  1. other [Section V.I (Real Agent Implementation Validation); Eq. (6); Section V.C metric definition]
    "The validation utilized three contractor types: GPT-4-Real ($2.00/task), Claude-3-Real ($1.50/task), and Budget-Cloud-Real ($0.80/task), competing against local execution costs of $0.00002/task. ... With proper epsilon-greedy exploration (10% rate), performance improved to 20.3% cost reduction with 11.4% outsourcing rate, demonstrating that exploration enables discovery of beneficial contractor relationships while maintaining economic rationality."

    Under Eq. (6), C_external(T,A_j) = P_j(T) + C_communication + C_verification + C_integration + C_risk + C_latency_penalty. Every additive term is nonnegative and the cheapest contractor price P_j is $0.80/task, so any outsourced task costs at least $0.80, while C_internal is stated as $0.00002/task. The decision algorithm would therefore always choose LOCAL, yet the paper reports 11.4% outsourcing and 20.3% cost reduction (and even 1.9% without exploration). Those numbers cannot be computed from the specified equations; they are imported as an assertion, so the validation is equivalent to assuming the result rather than deriving it.

  2. fitted input called prediction [Section IV.A Eq. (19); Algorithm 2 (Listing 2); Section V.I]
    "P(exploration) = ε = 0.1 (19) where ε represents the exploration rate, calibrated to 10% based on empirical analysis across diverse operational scenarios. ... With proper epsilon-greedy exploration (10% rate), performance improved to 20.3% cost reduction with 11.4% outsourcing rate, demonstrating that exploration enables discovery of beneficial contractor relationships while maintaining economic rationality."

    Algorithm 2 returns OUTSOURCE unconditionally whenever random() < ε, so the reported 11.4% outsourcing rate in the exploration run is essentially the pre-chosen ε=0.1 materializing as forced outsources. The '20.3% vs 1.9%' contrast is generated by switching this calibrated parameter on or off, and the savings attributed to those forced outsources are scored by the same unstated cost accounting that contradicts Eq. (6). The finding that ε-greedy exploration is essential is therefore a fitted parameter renamed as an empirical discovery.

full rationale

The paper's independent architectural content (hybrid skill representation, cost equations, A2A-based communication design) is not itself circular, and the self-citations [25], [27], [38]–[40] are background or architectural references rather than load-bearing validation. The central problem is in the validation chain. The 41.8% theoretical figure is a simulation-internal result: Section VI.A states that 'agent capabilities, costs, and performance metrics are modeled using carefully calibrated statistical distributions,' so it is an internal consistency check of the decision algorithm, not an externally anchored prediction; I do not count that alone as circular, since synthetic simulation is a legitimate way to probe algorithmic behavior. The decisive step is the real-agent validation: the stated local cost ($0.00002/task) and contractor prices ($0.80–$2.00/task) make every outsourcing decision strictly more expensive under Eq. (6), so the claimed 20.3% cost reduction and 344.4% per-task savings cannot be outputs of the paper's own cost model. That makes the headline empirical claim an imported assumption rather than a derived prediction, and the 'exploration is essential' conclusion is largely the calibrated ε forcing the outsourcing rate. This is partial circularity: the prediction reduces by construction to the chosen exploration parameter and to an unstated cost accounting, even though the framework proposal itself has independent content.

Assumptions & free parameters 13 free parameters · 5 assumptions · 0 invented entities

The paper's quantitative claims depend on many hand-set parameters (epsilon, thresholds, contractor prices, local-cost baseline) and on assumptions that synthetic market distributions represent real conditions. No code or data is shipped, so these choices cannot be audited. The $0.00002/task local baseline is especially load-bearing because the real-validation savings are computed against it.

free parameters (13)
  • Epsilon (exploration rate) = 0.1
    Calibrated by sensitivity analysis; the real-validation savings depend entirely on forcing 10% random outsourcing.
  • Local execution cost per task = $0.00002
    Baseline for the 240 real tasks; unverified and inconsistent with the claimed outsourcing savings.
  • GPT-4 contractor price = $2.00/task
    Assumed flat price for one real contractor type; drives external cost calculations.
  • Claude-3.5 contractor price = $1.50/task
    Assumed flat price in real validation.
  • Budget contractor price = $0.80/task
    Assumed flat price in real validation.
  • Theta_skill (skill compatibility threshold) = 0.7
    Hand-chosen filter; sensitivity analysis shows 0.5 and 0.9 degrade performance by 12.4% and 8.9%.
  • Tau_threshold (TOPSIS threshold) = 0.6
    Hand-chosen; determines when outsourcing beats local execution.
  • Rho_min (minimum confidence) = 0.8
    Hand-chosen; gates outsourcing decisions.
  • Eta (learning rate for weight updates) = 0.01
    Hand-chosen; affects dynamic weight adaptation.
  • Alpha (correlation penalty) = 0.3
    Chosen with a claimed 0.05 confidence interval from sensitivity analysis.
  • Beta (market responsiveness factor) = 0.7
    Hand-chosen; sensitivity analysis claims stability across 0.5-0.9.
  • EWMA lambda (cost calibration rate) = 0.1-0.3
    Range given without a specific value; affects cost estimates.
  • Skill matching weights (alpha, beta, gamma) = 0.3, 0.5, 0.2
    Typical values asserted in Section III.F.2.
assumptions (5)
  • domain assumption The simulated market prices, demand, and contractor capabilities are representative of real LLM agent markets.
    The simulation in Section V.A generates synthetic agent behavior from calibrated probability distributions; the 41.8% cost reduction depends on these distributions.
  • domain assumption Contractor agents truthfully advertise skills and prices and reliably execute accepted tasks.
    Section VI.B explicitly states the simulation assumes perfect contractor reliability and honest reporting, and that real ecosystems must handle malicious actors.
  • standard math The decision process can be modeled as a finite, irreducible, aperiodic Markov chain with a unique stationary distribution.
    Theorem 1 in Section III.F.8 invokes Perron-Frobenius but never specifies the transition matrix or verifies its spectral gap.
  • ad hoc to paper Local execution cost of $0.00002 per task is the correct baseline for the 240 real validation tasks.
    Section V.I asserts this value without token counts or cost breakdown; it is the baseline against which the 20.3% savings is computed.
  • domain assumption TOPSIS scores and correlation-adjusted weights are appropriate for comparing probabilistic LLM agent capabilities.
    Section III.F assumes MCDA is valid for qualitatively described and probabilistically executed LLM skills, an assumption the related-work section itself questions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of COALESCE: Economic and Security Dynamics of Skill-Based Task Outsourcing Among Team of Autonomous LLM Agents." pith.science (2026). https://pith.science/paper/EN5UQKC4

@misc{pith2026250601900,
  author       = {Pith},
  title        = {Pith review of: COALESCE: Economic and Security Dynamics of Skill-Based Task Outsourcing Among Team of Autonomous LLM Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EN5UQKC4}},
  note         = {Machine review of arXiv:2506.01900}
}
read the original abstract

The meteoric rise and proliferation of autonomous Large Language Model (LLM) agents promise significant capabilities across various domains. However, their deployment is increasingly constrained by substantial computational demands, specifically for Graphics Processing Unit (GPU) resources. This paper addresses the critical problem of optimizing resource utilization in LLM agent systems. We introduce COALESCE (Cost-Optimized and Secure Agent Labour Exchange via Skill-based Competence Estimation), a novel framework designed to enable autonomous LLM agents to dynamically outsource specific subtasks to specialized, cost-effective third-party LLM agents. The framework integrates mechanisms for hybrid skill representation, dynamic skill discovery, automated task decomposition, a unified cost model comparing internal execution costs against external outsourcing prices, simplified market-based decision-making algorithms, and a standardized communication protocol between LLM agents. Comprehensive validation through 239 theoretical simulations demonstrates 41.8\% cost reduction potential, while large-scale empirical validation across 240 real LLM tasks confirms 20.3\% cost reduction with proper epsilon-greedy exploration, establishing both theoretical viability and practical effectiveness. The emergence of proposed open standards like Google's Agent2Agent (A2A) protocol further underscores the need for frameworks like COALESCE that can leverage such standards for efficient agent interaction. By facilitating a dynamic market for agent capabilities, potentially utilizing protocols like A2A for communication, COALESCE aims to significantly reduce operational costs, enhance system scalability, and foster the emergence of specialized agent economies, making complex LLM agent functionalities more accessible and economically viable.

Figures

Figures reproduced from arXiv: 2506.01900 by the authors.

Figure 1
Figure 1. High-level workflow within COALESCE. Card JSON file, typically located at a well-known URL (/.well-known/agent.json) for a potential contrac￾tor agent. The client parses the card to determine the agent’s capabilities, endpoint, authentication needs, and supported protocols. This allows for standardized, direct discovery without necessarily relying on a central registry [40]. • Registry-Based Discovery: A central or … view at source ↗
Figure 2
Figure 2. Step-by-step Algorithmic Flow. input, it can transition the task state to input-required, prompting the Client to send subsequent messages for the same Task ID [10]. • Phase 4: Result Delivery (Contractor → Client via A2A): Upon completion, the Contractor’s A2A server transitions the task state to completed and sends the results back to the Client, typically packaged as an A2A Artifact containing relevant output Par… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Agent Capability Negotiation and Binding Protocol (ACNBP)

    cs.AI 2025-06 reject novelty 4.0 of 10

    ACNBP is a proposed standard for secure agent capability negotiation with an extension mechanism, but it lacks formal verification, experiments, and independent evaluation.

Reference graph

Works this paper leans on

44 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    The rise and potential of large language model based agents: A survey,

    Z. Xi, W. Chen, X. Guo, W. He, Y . Ding, B. Hong, M. Zhang, J. Wang, S. Jin, E. Zhouet al., “The rise and potential of large language model based agents: A survey,”arXiv preprint arXiv:2309.07864, 2023. [Online]. Available: https://arxiv.org/abs/2309.07864

  2. [2]

    Why Do Multi-Agent LLM Systems Fail? A Comprehensive Taxonomy, Dataset, and Mitigation Strategies,

    M. Cemri, R. Zhang, H. Yang, J. Li, C. Wang, L. Liu, and B. Liu, “Why Do Multi-Agent LLM Systems Fail? A Comprehensive Taxonomy, Dataset, and Mitigation Strategies,”arXiv preprint arXiv:2503.13657, Mar. 2025. [Online]. Available: https://arxiv.org/abs/2503.13657

  3. [3]

    LLM inference optimization,

    Hugging Face, “LLM inference optimization,” Transformers Documentation. [Online]. Available: https://huggingface.co/docs/ transformers/v4.35.2/llm_tutorial_optimization

  4. [4]

    GPU Memory Requirements For LLMs: Calculating VRAM,

    UnfoldAI Blog, “GPU Memory Requirements For LLMs: Calculating VRAM,” UnfoldAI, 2024. [Online]. Available: https://unfoldai.com/ gpu-memory-requirements-for-llms/

  5. [5]

    Retrieval-augmented generation for knowledge-intensive NLP tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, S. Riedel, and D. Kiela, “Retrieval-augmented generation for knowledge-intensive NLP tasks,” in Advances in Neural Information Processing Systems, vol. 33, 2020, pp. 9459–9474

  6. [6]

    Mastering LLM Techniques: Inference Optimization,

    NVIDIA Developer Blog, “Mastering LLM Techniques: Inference Optimization,” NVIDIA, 2024. [Online]. Available: https://developer. nvidia.com/blog/mastering-llm-techniques-inference-optimization/

  7. [7]

    Distributed systems meet economics: Pricing in the cloud,

    H. Wang, B. Li, and K. Li, “Distributed systems meet economics: Pricing in the cloud,” inProc. 2nd USENIX Workshop Hot Topics Cloud Comput. (HotCloud ’10), 2010. [Online]. Available: https: //www.usenix.org/legacy/event/hotcloud10/tech/full_papers/WangH.pdf

  8. [8]

    Navigating the High Cost of AI Compute,

    A. B. Downey, “Navigating the High Cost of AI Compute,” Andreessen Horowitz Blog, Apr. 2023. [Online]. Available: https: //a16z.com/navigating-the-high-cost-of-ai-compute/

Show all 44 references
  1. [9]

    The Cost of Computing,

    J. Gray, “The Cost of Computing,” Microsoft Research, Tech. Rep. MSR-TR-2003-24, 2003. [Online]. Available: https://www.microsoft. com/en-us/research/wp-content/uploads/2016/02/tr-2003-24.pdf

  2. [10]

    A2A: A new era of agent interoperability,

    Google Developers Blog, “A2A: A new era of agent interoperability,” Google, 2025, see also: Google, "Agent2Agent Protocol Docu- mentation," https://google.github.io/A2A/. [Online]. Available: https:// developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/

  3. [11]

    ALMA: Hierarchical learning for composite multi-agent tasks,

    S. Iqbal, R. Bapuraj, and F. Sha, “ALMA: Hierarchical learning for composite multi-agent tasks,” inAdvances in Neural Information Processing Systems, vol. 35, 2022. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/ 2022/file/2f27964513a28d034530bfdd117ea31...

  4. [12]

    Agentpoison: Red- teaming llm agents via poisoning memory or knowledge bases,

    Z. Chen, Z. Xiang, C. Xiao, D. Song, and B. Li, “Agentpoison: Red- teaming llm agents via poisoning memory or knowledge bases,”Ad- vances in Neural Information Processing Systems, vol. 37, pp. 130 185– 130 213, 2024

  5. [13]

    Multi-Agent Systems: A Survey,

    A. Dorri, S. S. Kanhere, and R. Jurdak, “Multi-Agent Systems: A Survey,”IEEE Access, vol. 6, pp. 28 573–28 593, 2018

  6. [14]

    Multi-agent deep reinforcement learning: a survey,

    S. Gronauer and K. Diepold, “Multi-agent deep reinforcement learning: a survey,”Artif. Intell. Rev., vol. 55, no. 2, pp. 895–943, 2022

  7. [15]

    A formal analysis and taxonomy of task allocation in multi-robot systems,

    B. P. Gerkey and M. J. Matari ´c, “A formal analysis and taxonomy of task allocation in multi-robot systems,”Int. J. Robot. Res., vol. 23, no. 9, pp. 939–954, 2004

  8. [16]

    The contract net protocol: High-level communication and control in a distributed problem solver,

    R. G. Smith, “The contract net protocol: High-level communication and control in a distributed problem solver,”IEEE Trans. Comput., vol. C-29, no. 12, pp. 1104–1113, Dec. 1980

  9. [17]

    Various Standards

    Foundation for Intelligent Physical Agents (FIPA), “Various Standards.” [Online]. Available: http://www.fipa.org/

  10. [18]

    An implementation of the contract net protocol based on marginal cost calculations,

    T. Sandholm, “An implementation of the contract net protocol based on marginal cost calculations,” inProc. 11th Natl. Conf. Artif. Intell. (AAAI-93), 1993, pp. 256–262

  11. [19]

    A Two-Stage Distributed Task Assignment Algorithm Based on Contract Net Protocol for Multi-UA V Cooperative Reconnaissance Task Reassignment in Dynamic Environ- ments,

    Y . Liu, H. Chen, Z. Li, and H. Wang, “A Two-Stage Distributed Task Assignment Algorithm Based on Contract Net Protocol for Multi-UA V Cooperative Reconnaissance Task Reassignment in Dynamic Environ- ments,”Sensors, vol. 23, no. 18, p. 7980, Sep. 2023

  12. [20]

    Double auctions in markets for multiple indivisible goods with budget constraints,

    J. Jantschgi, T. Heinrich, and A. Kaul, “Double auctions in markets for multiple indivisible goods with budget constraints,”Games and Economic Behavior, vol. 144, pp. 166–195, 2024

  13. [21]

    Learning in Double Auctions with Two-Sided Private Information,

    J. C. Jamison and B. Sundararajan, “Learning in Double Auctions with Two-Sided Private Information,” inProc. 39th Int. Conf. Mach. Learn. (ICML), 2022. [Online]. Available: https://openreview.net/pdf? id=2nTpPxJ5Bs

  14. [22]

    Computing VCG payments in combina- torial auctions: An overview,

    V . Conitzer and T. Sandholm, “Computing VCG payments in combina- torial auctions: An overview,” inProc. AAAI Workshop Auction Mech. E-Commerce, 2006

  15. [23]

    Types of Auctions and Their Properties,

    Fiveable Library, “Types of Auctions and Their Properties,”

  16. [24]

    Efficient mechanisms for bilateral trading,

    R. B. Myerson and M. A. Satterthwaite, “Efficient mechanisms for bilateral trading,”J. Econ. Theory, vol. 29, no. 2, pp. 265–281, 1983

  17. [25]

    Enterprise-Grade Security for the Model Context Protocol (MCP): Frameworks and Mitigation Strategies,

    V . S. Narajala and I. Habler, “Enterprise-Grade Security for the Model Context Protocol (MCP): Frameworks and Mitigation Strategies,”arXiv preprint arXiv:2504.08623, 2025. [Online]. Available: https://arxiv.org/abs/2504.08623

  18. [26]

    A Tutorial on Linear Function Approximators for Dynamic Programming and Reinforcement Learning,

    A. Geramifard, C. Dann, B. Kveton, B. Boots, and D. Fox, “A Tutorial on Linear Function Approximators for Dynamic Programming and Reinforcement Learning,”Foundations and Trends in Machine Learning, vol. 6, no. 4, pp. 375–451, 2013

  19. [27]

    Multi-agentic system threat modelling guide OW ASP GenAI security project,

    K. Huang, A. Sheriff, J. Sotiropoulos, R. F. Del, and V . Lu, “Multi-agentic system threat modelling guide OW ASP GenAI security project,” Apr. 2025. [Online]. Available: https://www.researchgate.net/publication/391204915_Multi-Agentic_ system_Threat_Modelling_Guide_OW ASP_Gen...

  20. [28]

    Ontologies for Multi-Agent Systems Specification,

    K. K. Breitman and J. C. S. P. Leite, “Ontologies for Multi-Agent Systems Specification,” NASA Technical Reports Server (NTRS), Tech. Rep., 2005. [Online]. Available: https://ntrs.nasa.gov/citations/ 20050137693

  21. [29]

    Rule-based systems,

    G. Governatori and A. Rotolo, “Rule-based systems,” inHandbook of Research on Discrete Event Simulation Environments: Technologies and Applications. IGI Global, 2010, pp. 408–431

  22. [30]

    Hierarchical Multi-Agent Skill Discovery,

    Z. Li, L. Chen, J. Liu, W. Zhang, and D. Zhao, “Hierarchical Multi-Agent Skill Discovery,” inAdvances in Neural Information Processing Systems, vol. 36, 2023. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/ 2023/file/c276c3303c0723c83a43b95a44a1fcbf-Pape...

  23. [31]

    Transaction costs of cloud computing: A 360-degree industry analysis,

    M. Makhlouf, “Transaction costs of cloud computing: A 360-degree industry analysis,”J. Cloud Comput.: Adv., Syst. Appl., vol. 9, no. 1,

  24. [32]

    Economic models for resource management and scheduling in Grid computing,

    R. Buyya, D. Abramson, J. Giddy, and H. Stockinger, “Economic models for resource management and scheduling in Grid computing,” Concurrency Computat.: Pract. Exper., vol. 14, no. 13-15, pp. 1507– 1542, 2002

  25. [33]

    Agent-Based Computational Economics (ACE): Overview,

    L. Tesfatsion, “Agent-Based Computational Economics (ACE): Overview,” Iowa State University Faculty Site, n.d. [Online]. Available: https://faculty.sites.iastate.edu/tesfatsi/archive/tesfatsi/ace.htm

  26. [34]

    Cloud Computing: Competition Issues,

    M. Bourreau, C. Caffarra, E. Carroni, and F. S. Morton, “Cloud Computing: Competition Issues,” TSE Working Paper, Tech. Rep. n. 24-1520, 2024. [Online]. Available: https://www.tse-fr.eu/sites/default/ files/TSE/documents/doc/wp/2024/wp_tse_1520.pdf

  27. [35]

    Understanding the planning of LLM agents: A survey,

    X. Huang, H. Kwak, and J. An, “Understanding the planning of LLM agents: A survey,”arXiv preprint arXiv:2402.02716, Feb. 2024. [Online]. Available: https://arxiv.org/abs/2402.02716

  28. [36]

    TwoStep: Multi-agent Task Planning using Classical Planners and Large Language Models,

    D. Bai, I. Singh, D. Traum, and J. Thomason, “TwoStep: Multi-agent Task Planning using Classical Planners and Large Language Models,” arXiv preprint arXiv:2403.17246, Mar. 2024. [Online]. Available: https://arxiv.org/abs/2403.17246

  29. [37]

    TDAG: A Multi-Agent Framework based on Dynamic Task Decomposition and Agent Generation,

    Z. Liu, K. Liu, Z. Yuan, C. Wang, Z. Chen, K. Wang, K. Yuan, K. Chen, and W. X. Zhao, “TDAG: A Multi-Agent Framework based on Dynamic Task Decomposition and Agent Generation,” arXiv preprint arXiv:2402.10178, Feb. 2024. [Online]. Available: https://arxiv.org/abs/2402.10178

  30. [38]

    Agent name service (ans): A universal directory for secure ai agent discovery and interoperability,

    H. Ken, V . S. Narajala, I. Habler, and A. Sheriff, “Agent name service (ans): A universal directory for secure ai agent discovery and interoperability,”arXiv preprint arXiv:2505.10609, 2025. [Online]. Available: https://arxiv.org/abs/2505.10609

  31. [39]

    Building a secure agentic AI application leveraging A2A protocol,

    I. Habler, K. Huang, V . S. Narajala, and P. Kulkarni, “Building a secure agentic AI application leveraging A2A protocol,” 2025. [Online]. Available: https://www.arxiv.org/abs/2504.16902

  32. [40]

    Securing genai multi- agent systems against tool squatting: A zero trust registry-based approach,

    V . S. Narajala, K. Huang, and I. Habler, “Securing genai multi- agent systems against tool squatting: A zero trust registry-based approach,”arXiv preprint arXiv:2504.19951, 2025. [Online]. Available: https://arxiv.org/abs/2504.19951

  33. [41]

    Privacy Notice,

    UKG, “Privacy Notice,” n.d. [Online]. Available: https://www.ukg.com/ privacy

  34. [42]

    R. S. Sutton and A. G. Barto,Reinforcement learning: An introduction, 2nd ed. MIT press, 2018

  35. [2020]

    Available: https://d-nb.info/1208575686/34

    [Online]. Available: https://d-nb.info/1208575686/34

  36. [2024]

    Available: https://library.fiveable.me/game-theory/ unit-10/types-auctions-properties/study-guide/yuYCGjcRzDPLXG79

    [Online]. Available: https://library.fiveable.me/game-theory/ unit-10/types-auctions-properties/study-guide/yuYCGjcRzDPLXG79

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.