pith. sign in

arxiv: 2605.28764 · v1 · pith:BYRO4QXUnew · submitted 2026-05-27 · 💻 cs.AI · cs.DC· cs.MA

SwarmHarness: Skill-Based Task Routing via Decentralized Incentive-Aligned AI Agent Networks

Edwin Jose This is my paper

Pith reviewed 2026-06-29 11:39 UTC · model grok-4.3

classification 💻 cs.AI cs.DCcs.MA
keywords decentralized computeincentive mechanismsswarm intelligencetask routingAI agent networksShapley value approximationdistributed hash tableself-organizing systems
0
0 comments X

The pith

SwarmHarness lets nodes self-organize into compute swarms where credits and skill-based routing produce emergent collective intelligence without central control.

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

The paper introduces SwarmHarness as a protocol that combines a distributed hash table for node discovery, a utility-based router for task assignment, and a credit system for incentives to enable safe sharing of unused compute resources. Nodes earn credits by completing tasks and spend them to submit work, while idle nodes lose priority, aiming to create self-regulation. As nodes specialize in rewarded skills and routing acts like digital pheromones, the network is claimed to develop swarm-like collective intelligence. This matters because it targets the problem of wasted GPU cycles on personal devices and edge hardware that current centralized or blockchain-heavy systems fail to unlock efficiently. The result would be a foundation for autonomous AI agent networks that allocate and settle compute independently.

Core claim

SwarmHarness demonstrates that a DHT-based SwarmRegistry for capability advertisement, a SwarmRouter using utility over capability load latency and trust, and a SwarmCredit mechanism with Shapley-value reward attribution allow nodes to self-organize into a compute swarm. Nodes earn credits for serving tasks and spend them to submit, with idle nodes draining credits and losing priority, producing a self-regulating participation economy. As specialization occurs toward high-reward skills and routing signals function as digital pheromones, the network exhibits emergent collective intelligence analogous to biological swarms, extending beyond compute sharing to autonomous distributed AI agent net

What carries the argument

SwarmCredit, the incentive layer that attributes rewards via Shapley-value approximation so that nodes earn credits for serving tasks, spend them to submit tasks, and lose priority when idle, thereby creating self-regulation and driving skill specialization.

Load-bearing premise

The credit mechanism will automatically produce balanced participation and prevent free-riding or collapse without any central authority or extra rules.

What would settle it

A live deployment or large-scale simulation in which many nodes remain idle yet retain high routing priority and the overall participation level drops below the threshold needed for task completion.

Figures

Figures reproduced from arXiv: 2605.28764 by Edwin Jose.

Figure 1
Figure 1. Figure 1: SwarmHarness end-to-end architecture. A user submits a task to the SwarmRouter, which queries the [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
read the original abstract

Vast quantities of compute (GPU cycles on personal workstations, idle inference servers, and edge devices between jobs) go unused because no incentive-aligned protocol exists for their owners to share them safely and profitably. Existing approaches either require a trusted central coordinator (cloud marketplaces), demand heavy blockchain infrastructure (Golem, BrokerChain), or lack an incentive layer entirely (BOINC, Petals). We propose SwarmHarness, a decentralised protocol in which HarnessAPI skill nodes self-organise into a compute swarm without any central authority. SwarmHarness has three interlocking components: a SwarmRegistry built on a Distributed Hash Table (DHT) for peer discovery and capability advertisement; a SwarmRouter that dispatches tasks to nodes using a utility function over capability, load, latency, and trust; and SwarmCredit, an incentive mechanism that attributes compute-credit rewards to contributing nodes via a Shapley-value approximation. Nodes earn credits by serving tasks and spend credits to submit them; idle nodes that never contribute drain credits and lose routing priority, creating a self-regulating participation economy. As nodes specialise toward high-reward skills and routing signals act as digital pheromones, the network exhibits emergent collective intelligence analogous to biological swarms. Beyond compute sharing, SwarmHarness is a foundational primitive for autonomous distributed AI agent networks in which agents hire compute, route subtasks, and settle credits without human intermediation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes SwarmHarness, a decentralized protocol for sharing unused compute (GPUs, edge devices) among AI agents without central authority or heavy blockchain. It outlines three components: a DHT-based SwarmRegistry for peer discovery and capability advertisement; a SwarmRouter using a utility function over capability, load, latency, and trust for task dispatch; and SwarmCredit, an incentive layer that attributes rewards via Shapley-value approximation, with nodes earning credits by serving tasks, spending them to submit tasks, and losing priority when idle. The central claim is that these mechanisms produce self-regulating participation, skill specialization, routing signals that act as digital pheromones, and emergent collective intelligence analogous to biological swarms, enabling autonomous agent networks.

Significance. If the incentive and routing dynamics were validated to produce stable specialization and collective behavior, the proposal could provide a useful primitive for decentralized compute markets and multi-agent systems, filling gaps left by centralized marketplaces, blockchain-heavy systems like Golem, or incentive-free projects like BOINC and Petals.

major comments (3)
  1. [Abstract] Abstract (SwarmCredit paragraph): the claim that the earn/spend/idle-drain rules 'create a self-regulating participation economy' and prevent collapse into low-participation states is presented without equilibrium analysis, game-theoretic modeling of free-riding or collusion, differential-equation or agent-based dynamics, or any simulation results.
  2. [Abstract] Abstract (final sentence): the assertion that 'nodes specialise toward high-reward skills and routing signals act as digital pheromones' yielding 'emergent collective intelligence analogous to biological swarms' rests entirely on the unexamined component descriptions and lacks any formalization, convergence argument, or empirical support.
  3. [Abstract] Abstract (SwarmCredit description): the Shapley-value approximation for credit attribution is invoked without pseudocode, complexity bounds, or incentive-compatibility arguments showing it is robust to strategic misreporting.
minor comments (1)
  1. The manuscript would benefit from explicit comparison tables or diagrams contrasting SwarmHarness components against Golem, BrokerChain, BOINC, and Petals.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for identifying the unsubstantiated claims in the abstract. The manuscript is a conceptual protocol proposal rather than an empirical or theoretical analysis paper; we will revise the abstract and add qualifying language to avoid overstating results while preserving the design rationale.

read point-by-point responses

  1. Referee: [Abstract] Abstract (SwarmCredit paragraph): the claim that the earn/spend/idle-drain rules 'create a self-regulating participation economy' and prevent collapse into low-participation states is presented without equilibrium analysis, game-theoretic modeling of free-riding or collusion, differential-equation or agent-based dynamics, or any simulation results.

    Authors: We agree the abstract presents this outcome as achieved rather than conjectured. The design intends the credit rules to discourage free-riding via priority loss, but no equilibrium analysis or simulations are present. We will revise the abstract to state that the rules are designed to produce self-regulation and note this as a hypothesis for future validation. revision: yes

  2. Referee: [Abstract] Abstract (final sentence): the assertion that 'nodes specialise toward high-reward skills and routing signals act as digital pheromones' yielding 'emergent collective intelligence analogous to biological swarms' rests entirely on the unexamined component descriptions and lacks any formalization, convergence argument, or empirical support.

    Authors: The analogy is drawn from the intended dynamics of specialization and pheromone-like routing, but we acknowledge it lacks formalization or evidence. We will revise the final sentence to present these as expected behaviors of the protocol design rather than demonstrated emergent properties. revision: yes

  3. Referee: [Abstract] Abstract (SwarmCredit description): the Shapley-value approximation for credit attribution is invoked without pseudocode, complexity bounds, or incentive-compatibility arguments showing it is robust to strategic misreporting.

    Authors: The manuscript describes the approximation at a conceptual level only. We will add a brief methods subsection with pseudocode for the approximation and a short discussion of its computational properties and basic incentive alignment, while noting that a full strategic analysis lies outside the current scope. revision: partial

Circularity Check

0 steps flagged

No circularity: conceptual proposal without derivations or predictions

full rationale

The manuscript is a high-level protocol proposal describing SwarmRegistry (DHT), SwarmRouter (utility function), and SwarmCredit (credit earn/spend with idle drain) at the level of component names and intended behaviors. No equations, fitted parameters, predictions, or derivation steps appear in the text. The emergent-intelligence claim is presented as an expected outcome of the rules rather than the result of any chain that reduces to its own inputs by construction. Because no analytical derivation exists to inspect, the circularity score is 0.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The proposal relies on standard assumptions from distributed systems and mechanism design but introduces no new formal entities or parameters in the abstract.

axioms (1)
  • domain assumption Shapley-value approximation is feasible and incentive-compatible for credit attribution in this setting
    Invoked for the SwarmCredit component without further justification in the abstract.

pith-pipeline@v0.9.1-grok · 5775 in / 1245 out tokens · 32727 ms · 2026-06-29T11:39:14.520071+00:00 · methodology

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

28 extracted references · 22 canonical work pages · 2 internal anchors

  1. [1]

    HarnessAPI: A skill-first framework for unified streaming APIs and MCP tools

    Edwin Jose. HarnessAPI: A skill-first framework for unified streaming APIs and MCP tools. arXiv preprint, 2026. arXiv preprint, May 2026

    2026

  2. [2]

    BOINC: A Platform for Volunteer Computing

    David P. Anderson. BOINC: A platform for volunteer computing. arXiv:1903.01699, 2019. arXiv:1903.01699

    work page internal anchor Pith review Pith/arXiv arXiv 1903

  3. [3]

    Petals: Collaborative inference and fine-tuning of large models

    Alexander Borzunov, Dmitry Baranchuk, Tim Dettmers, Max Ryabinin, Younes Belkada, Artem Chu- machenko, Pavel Samygin, and Colin Raffel. Petals: Collaborative inference and fine-tuning of large models. arXiv:2209.01188, 2022. arXiv:2209.01188

    work page arXiv 2022

  4. [4]

    Communication-efficient learning of deep networks from decentralized data,

    H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication- efficient learning of deep networks from decentralized data. arXiv:1602.05629, 2023. arXiv:1602.05629

    work page arXiv 2023

  5. [5]

    Golem: A decentralised supercomputer

    Golem Factory. Golem: A decentralised supercomputer. arXiv:1801.04024, 2018. arXiv:1801.04024; Technical whitepaper

    work page arXiv 2018

  6. [6]

    Decentralized AI infrastructure built on BrokerChain

    Jianru Lin. Decentralized AI infrastructure built on BrokerChain. InProceedings of the ACM, 2025

    2025

  7. [7]

    ReAct: Synergizing Reasoning and Acting in Language Models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. arXiv:2210.03629, 2022. arXiv:2210.03629

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  8. [8]

    AgentMesh: A cooperative multi-agent generative AI framework for software development automation

    Milad Khanzadeh. AgentMesh: A cooperative multi-agent generative AI framework for software development automation. arXiv:2507.19902, 2025. arXiv:2507.19902

    work page arXiv 2025

  9. [9]

    Harnessing multi-agent LLMs for complex engineering problem-solving: A swarm intelligence approach

    Mona Almasoud et al. Harnessing multi-agent LLMs for complex engineering problem-solving: A swarm intelligence approach. arXiv:2501.01205, 2025. arXiv:2501.01205; IEEE EDUCON 2025

    work page arXiv 2025

  10. [10]

    Distributed deep learning using volunteer computing-like paradigm

    Madhurima Atre, Bhavya Jha, and Yogesh Simmhan. Distributed deep learning using volunteer computing-like paradigm. arXiv:2103.08894, 2021. arXiv:2103.08894; IEEE IPDPSW 2021

    work page arXiv 2021

  11. [11]

    Lattica: A decentralized cross-NAT communication framework for scalable AI inference and training

    Yang et al. Lattica: A decentralized cross-NAT communication framework for scalable AI inference and training. arXiv:2510.00183, 2025. arXiv:2510.00183

    work page arXiv 2025

  12. [12]

    FedToken: Tokenized incentives for data contribution in federated learning

    Shashi Raj Pandey, Lam Duc Nguyen, and Petar Popovski. FedToken: Tokenized incentives for data contribution in federated learning. arXiv:2209.09775, 2022. arXiv:2209.09775

    work page arXiv 2022

  13. [13]

    A contract theory based incentive mechanism for federated learning

    Mengmeng Tian, Yuxin Chen, Yuan Liu, Zehui Xiong, Cyril Leung, and Chunyan Miao. A contract theory based incentive mechanism for federated learning. arXiv:2108.05568, 2021. arXiv:2108.05568

    work page arXiv 2021

  14. [14]

    Right reward right time for federated learning

    Nguyen et al. Right reward right time for federated learning. arXiv:2503.07869, 2025. arXiv:2503.07869

    work page arXiv 2025

  15. [15]

    Chanuka A. S. Hewa Kaluannakkage and Rajkumar Buyya. Incentive-based federated learning: Architectural elements and future directions. arXiv:2510.14208, 2025. arXiv:2510.14208

    work page arXiv 2025

  16. [16]

    Friedlander, and Yong Zhang

    Zhenan Fan, Huang Fang, Xinglu Wang, Zirui Zhou, Jian Pei, Michael P. Friedlander, and Yong Zhang. Fair and efficient contribution valuation for vertical federated learning. arXiv:2201.02658, 2022. arXiv:2201.02658

    work page arXiv 2022

  17. [17]

    A survey of multi-agent deep reinforcement learning with communication

    Changxi Zhu, Mehdi Dastani, and Shihan Wang. A survey of multi-agent deep reinforcement learning with communication. arXiv:2203.08975, 2022. arXiv:2203.08975

    work page arXiv 2022

  18. [18]

    An introduction to centralized training for decentralized execution in cooperative multi-agent re- inforcement learning.arXiv preprint arXiv:2409.03052,

    Christopher Amato. An introduction to centralized training for decentralized execution in cooperative multi-agent reinforcement learning. arXiv:2409.03052, 2024. arXiv:2409.03052

    work page arXiv 2024

  19. [19]

    From pheromones to policies: Reinforcement learning for engineered biological swarms

    Vellinger, Antonic, and Tuci. From pheromones to policies: Reinforcement learning for engineered biological swarms. arXiv:2509.20095, 2025. arXiv:2509.20095

    work page arXiv 2025

  20. [20]

    Evolving neural networks reveal emergent collective behavior from minimal agent interactions

    Giardini, Hardy, and da Cunha. Evolving neural networks reveal emergent collective behavior from minimal agent interactions. arXiv:2410.19718, 2024. arXiv:2410.19718

    work page arXiv 2024

  21. [21]

    Internet of intelligence: Collective advantage for communications

    Li et al. Internet of intelligence: Collective advantage for communications. arXiv: 1905.00719, 2020. arXiv:1905.00719. 11 SwarmHarness: Skill-Based Task Routing via Decentralized Incentive-Aligned AI Agent Networks

    work page arXiv 1905

  22. [22]

    Communication-efficient training workload balancing for decentralized multi-agent learning

    Seyed Mahmoud Sajjadi Mohammadabadi, Lei Yang, Feng Yan, and Junshan Zhang. Communication-efficient training workload balancing for decentralized multi-agent learning. arXiv:2405.00839, 2024. arXiv:2405.00839; IEEE ICDCS 2024

    work page arXiv 2024

  23. [23]

    Blockchain-based federated learning: Incentivizing data sharing and penalizing dishonest behavior

    Jaberzadeh et al. Blockchain-based federated learning: Incentivizing data sharing and penalizing dishonest behavior. arXiv:2307.10492, 2023. arXiv:2307.10492

    work page arXiv 2023

  24. [24]

    The shapley value in machine learning

    Benedek Rozemberczki, Lauren Watson, Péter Bayer, Hao-Tsung Yang, Olivér Kiss, Sebastian Nilsson, and Rik Sarkar. The shapley value in machine learning. arXiv:2202.05594, 2022. arXiv:2202.05594

    work page arXiv 2022

  25. [25]

    Kademlia: A peer-to-peer information system based on the XOR metric

    Petar Maymounkov and David Mazières. Kademlia: A peer-to-peer information system based on the XOR metric. InRevised Papers from the 1st International Workshop on Peer-to-Peer Systems (IPTPS), pages 53–65. Springer, 2002

    2002

  26. [26]

    Gossip-based aggregation in large dynamic networks

    Márk Jelasity, Alberto Montresor, and Ozalp Babaoglu. Gossip-based aggregation in large dynamic networks. ACM Transactions on Computer Systems, 23(3):219–252, 2005

    2005

  27. [27]

    Bitcoin: A peer-to-peer electronic cash system

    Satoshi Nakamoto. Bitcoin: A peer-to-peer electronic cash system. https://bitcoin.org/bitcoin.pdf,

  28. [28]

    Accessed: May 2026. 12

    2026