pith. sign in

arxiv: 2606.30986 · v1 · pith:WL7WAUCNnew · submitted 2026-06-29 · 💻 cs.CY · cs.HC· cs.MA· econ.GN· q-fin.EC

The Organizational Behavior of Agentic AI: Collective Intelligence in Human-Agent Workflows

Canhui Liu This is my paper

Pith reviewed 2026-07-01 00:43 UTC · model grok-4.3

classification 💻 cs.CY cs.HCcs.MAecon.GNq-fin.EC
keywords agentic AIorganizational behaviorcollective intelligencecontextual transaction costhuman-agent workflowsLLM agentsworkflow coordination
0
0 comments X

The pith

Agentic AI collectives resemble human organizations in differentiating work and coordinating tasks but sustain these patterns through context architecture rather than motivation or trust.

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

The paper examines agentic AI systems deployed as collectives of planners, solvers, reviewers, memory managers, tool users, and orchestrators entering workflows labeled as teams or committees. It claims these collectives exhibit organizational behavior by differentiating work, coordinating interdependence, performing recurrent routines, crossing boundaries, and producing collective outcomes. The resemblance holds in structure but not in sustaining forces, which rely on prompts, memory, traces, schemas, tools, validators, and permissions instead of identity, trust, employment, or moral accountability. Contextual transaction cost is developed as the mechanism explaining when such systems succeed or fail. Simulations and real agent traces demonstrate that forms imitating human handoffs often underperform while shared-state designs improve when context remains durable and inspectable.

Core claim

Agentic artificial intelligence is a partial organisational analogue. It resembles a human organisation because it differentiates work, coordinates interdependence, performs recurrent routines, crosses boundaries, and produces collective outcomes. It differs because these patterns are not sustained by motivation, identity, trust, employment, socialisation, or moral accountability. They are sustained by context architecture: prompts, memory, traces, schemas, tools, validators, and permissions. The article develops contextual transaction cost as the central mechanism linking these similarities and differences. Computational theorising, synthetic task simulations, real LLM agent traces, and rob

What carries the argument

Contextual transaction cost, the mechanism that links structural similarities to performance differences by measuring the costs of maintaining and transferring context across agents in collectives.

If this is right

  • Human-imitation agent structures often add lossy handoffs, correlated deliberation, and verification burdens that reduce performance.
  • Shared-state and adaptive coordination forms perform better when context is made durable, inspectable, and task-contingent.
  • Agent collectives can jointly support collective intelligence with human workflows under interface conditions that align context architecture with task demands.
  • The patterns of work differentiation and coordination in agent systems can be analyzed and improved independently of social or motivational factors.

Where Pith is reading between the lines

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

  • Workflow designers should prioritize minimizing contextual transaction costs over directly copying human team structures when building multi-agent systems.
  • This view could extend to diagnosing coordination failures in existing agent deployments where context is lost or duplicated across steps.
  • Hybrid human-agent setups may gain from exposing agent memory and traces to human participants to lower overall coordination costs.
  • Testing on tasks with high interdependence could reveal whether durable context reduces error rates more than adding reviewer agents.

Load-bearing premise

Computational theorising, synthetic task simulations, and real LLM agent traces can demonstrate organisational behaviour and performance differences in a manner analytically comparable to human organisations despite the complete absence of social and motivational elements.

What would settle it

A controlled comparison in which human-imitation agent teams with high verification burdens achieve higher collective task performance than shared-state designs with durable context on identical interdependent tasks.

Figures

Figures reproduced from arXiv: 2606.30986 by Canhui Liu.

Figure 1
Figure 1. Figure 1: presents the theoretical architecture. Human workflows impose demands for judgment, accountability, continuity, and legitimacy. Agentic AI collectives generate differentiated and recur￾rent computational action. Their productive connection depends on interface organizations that translate between human accountability and agentic context coordination [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Organization form and contextual transaction cost in the simulation The family-level contrast is large. Agent-native forms are 395.26% more efficient than human￾imitation forms. Adaptive meta-organization improves collective efficiency by 89.24% relative to the single expert. Blackboard memory is 139.44% more efficient than the best human-imitation form. These differences are not simply artifacts of task a… view at source ↗
Figure 3
Figure 3. Figure 3: Efficient agent scale depends on communication structure and error correlation The interpretation is straightforward. Additional agents help when they contribute independent information, search paths, tools, or verification capacity. They hurt when they share the same error structure and communicate through lossy natural language. A five-agent system using the same model, prompt template, and retrieval con… view at source ↗
Figure 4
Figure 4. Figure 4: Real LLM trace evidence on quality and CTC-adjusted efficiency has the highest mean quality at 94.06 but lower efficiency at 45.11 because its CTC index rises to 9.79. Adaptive organization has mean quality of 92.61 and efficiency of 53.57. In the open￾source runs, single execution again has the highest efficiency because of low coordination cost, while adaptive and pipeline improve quality and success at … view at source ↗
read the original abstract

Agentic artificial intelligence is increasingly deployed not as a single assistant but as a collective of planners, solvers, reviewers, memory managers, tool users, and orchestrators. These systems are entering organisational workflows under familiar labels such as teams, managers, committees, markets, and workflows. This article asks whether such agent collectives exhibit organisational behaviour in a sense that is analytically comparable to, yet distinct from, human organisational behaviour. I argue that agentic AI is a partial organisational analogue. It resembles a human organisation because it differentiates work, coordinates interdependence, performs recurrent routines, crosses boundaries, and produces collective outcomes. It differs because these patterns are not sustained by motivation, identity, trust, employment, socialisation, or moral accountability. They are sustained by context architecture: prompts, memory, traces, schemas, tools, validators, and permissions. The article develops contextual transaction cost as the central mechanism linking these similarities and differences. Computational theorising, synthetic task simulations, real LLM agent traces, and robustness analyses show that human-imitation forms often underperform when they add lossy handoffs, correlated deliberation, and verification burdens, whereas shared-state and adaptive forms perform better when they make context durable, inspectable, and task-contingent. The article contributes to organisation studies by theorising agentic AI as an emerging object of organising and by specifying the interface conditions under which human and agentic organisational behaviour can jointly support collective intelligence.

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

2 major / 1 minor

Summary. The paper claims that agentic AI collectives exhibit organizational behaviour analogous to human organisations because they differentiate work, coordinate interdependence, perform recurrent routines, cross boundaries, and produce collective outcomes. It differs because these patterns are sustained by context architecture (prompts, memory, traces, schemas, tools, validators, permissions) rather than motivation, identity, trust, employment, socialisation, or moral accountability. The central mechanism is contextual transaction cost. Computational theorising, synthetic task simulations, real LLM agent traces, and robustness analyses show that human-imitation forms underperform due to lossy handoffs, correlated deliberation, and verification burdens, while shared-state and adaptive forms perform better when context is durable, inspectable, and task-contingent. The contribution is to organisation studies by theorising agentic AI as an object of organising and specifying interface conditions for joint human-agent collective intelligence.

Significance. If the results hold, the paper contributes to organisation studies by extending theory to agentic AI systems and providing a framework for collective intelligence in human-agent workflows. It introduces contextual transaction cost as a linking mechanism and uses simulations to compare organisational forms, which could inform AI system design and the study of emerging technologies. The use of synthetic tasks and LLM traces offers a computational approach to organisational questions.

major comments (2)
  1. [Simulations and traces section] Simulations and traces section: performance differences between human-imitation and shared-state forms are reported, but the organisational constructs (differentiation, interdependence coordination, routines, boundary crossing) are not operationalised or measured with established organisation-theory metrics or human baselines. This is load-bearing for the central claim of analytical comparability.
  2. [Contextual transaction cost development] Contextual transaction cost development: the construct is introduced and developed inside the paper to explain why context architecture substitutes for social mechanisms, yet simulation outcomes appear defined in terms of context durability and lossy handoffs, creating a risk of circularity rather than independent grounding.
minor comments (1)
  1. The abstract states that robustness analyses were performed but supplies no details on sample sizes, error measures, or exclusion rules; these should be added to the methods description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and the recommendation for major revision. The two major comments identify opportunities to strengthen the manuscript's links to organization theory. We respond point by point below.

read point-by-point responses

  1. Referee: [Simulations and traces section] Simulations and traces section: performance differences between human-imitation and shared-state forms are reported, but the organisational constructs (differentiation, interdependence coordination, routines, boundary crossing) are not operationalised or measured with established organisation-theory metrics or human baselines. This is load-bearing for the central claim of analytical comparability.

    Authors: We agree that greater explicitness would help. In revision we will add a dedicated subsection and mapping table that operationalizes each construct inside the simulation and trace designs: differentiation via role-specific prompts and tool permissions; interdependence coordination via shared-state versus message-passing protocols; routines via recurrent task sequences in the traces; and boundary crossing via validator and external-tool interfaces. Performance metrics (task completion, error propagation, coordination overhead) will be explicitly tied to these operationalizations. We will also clarify that human baselines fall outside the paper's scope, which compares alternative agent organizational forms rather than benchmarking against human teams; adding such baselines would require a separate empirical study. revision: partial

  2. Referee: [Contextual transaction cost development] Contextual transaction cost development: the construct is introduced and developed inside the paper to explain why context architecture substitutes for social mechanisms, yet simulation outcomes appear defined in terms of context durability and lossy handoffs, creating a risk of circularity rather than independent grounding.

    Authors: We disagree that the argument is circular. Contextual transaction cost is derived from classical transaction-cost economics and specified as the costs of context loss, duplication, and verification that arise when agents must hand off or reconstruct state. The simulations measure independent outcomes—task success rate, latency, and error types—that the theory predicts will differ across architectures. Descriptions of durability and handoffs are the observable mechanisms through which the costs operate, not the outcome variables themselves. In revision we will present the construct formally in a dedicated theory subsection before describing the simulation measures, making the separation between theoretical mechanism and empirical indicators explicit. revision: no

Circularity Check

0 steps flagged

No circularity; derivation self-contained with independent simulation evidence

full rationale

The abstract and description introduce contextual transaction cost as a new theoretical mechanism and present computational theorising plus synthetic task simulations and LLM agent traces as separate evidence showing performance differences (e.g., human-imitation forms underperform due to lossy handoffs). No equations, self-citations, fitted parameters renamed as predictions, or reductions of claims to inputs by construction are present in the provided text. The simulations are described as demonstrating outcomes independently, satisfying the requirement for external grounding rather than definitional equivalence.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on a domain-level analogy between human and agent collectives plus a newly introduced explanatory construct whose grounding is asserted via simulations whose details are absent from the abstract.

axioms (1)
  • domain assumption Organisational behaviour consists of work differentiation, interdependence coordination, recurrent routines, boundary crossing, and collective outcomes.
    Invoked to establish the basis for claiming resemblance between human organisations and agent collectives.
invented entities (1)
  • contextual transaction cost no independent evidence
    purpose: Central mechanism that links architectural similarities and differences to performance outcomes in agent collectives.
    Newly developed within the paper; no independent evidence outside the paper is described in the abstract.

pith-pipeline@v0.9.1-grok · 5794 in / 1593 out tokens · 54884 ms · 2026-07-01T00:43:54.547671+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

47 extracted references · 47 canonical work pages

  1. [1]

    A., & Demsetz, H

    Alchian, A. A., & Demsetz, H. (1972). Production, information costs, and economic organization. American Economic Review, 62(5), 777-795

    work page 1972

  2. [2]

    Arrow, K. J. (1974). The Limits of Organization. Norton

    work page 1974

  3. [3]

    Bai, Y., Lv, X., Zhang, J., Lyu, H., Tang, J., Huang, Z., Du, Z., Liu, X., Zeng, A., Hou, L., Dong, Y., Tang, J., & Li, J. (2024). LongBench: A bilingual, multitask benchmark for long context understanding. Proceedings of the Association for Computational Linguistics

    work page 2024

  4. [4]

    Barley, S. R. (1986). Technology as an occasion for structuring: Evidence from observations of CT scanners and the social order of radiology departments. Administrative Science Quarterly, 31(1), 78-108

    work page 1986

  5. [5]

    Benkler, Y. (2002). Coase’s Penguin, or, Linux and the nature of the firm. Yale Law Journal, 112(3), 369-446

    work page 2002

  6. [6]

    S., & Duguid, P

    Brown, J. S., & Duguid, P. (1991). Organizational learning and communities-of-practice: Toward a unified view of working, learning, and innovation. Organization Science, 2(1), 40-57

    work page 1991

  7. [7]

    Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889-942

    work page 2025

  8. [8]

    Carlile, P. R. (2002). A pragmatic view of knowledge and boundaries: Boundary objects in new product development. Organization Science, 13(4), 442-455

    work page 2002

  9. [9]

    Carlile, P. R. (2004). Transferring, translating, and transforming: An integrative framework for managing knowledge across boundaries. Organization Science, 15(5), 555-568

    work page 2004

  10. [10]

    Coase, R. H. (1937). The nature of the firm. Economica, 4(16), 386-405

    work page 1937

  11. [11]

    D., & Bacdayan, P

    Cohen, M. D., & Bacdayan, P. (1994). Organizational routines are stored as procedural memory: Evidence from a laboratory study. Organization Science, 5(4), 554-568

    work page 1994

  12. [12]

    M., & March, J

    Cyert, R. M., & March, J. G. (1963). A Behavioral Theory of the Firm. Prentice-Hall. 17

    work page 1963

  13. [13]

    A., & Gardner, M

    Dasigi, P., Lo, K., Beltagy, I., Cohan, A., Smith, N. A., & Gardner, M. (2021). A dataset of information-seeking questions and answers anchored in research papers. Proceedings of NAACL

    work page 2021

  14. [14]

    P., Eisenhardt, K

    Davis, J. P., Eisenhardt, K. M., & Bingham, C. B. (2007). Developing theory through simulation methods. Academy of Management Review, 32(2), 480-499

    work page 2007

  15. [15]

    Faraj, S., Pachidi, S., & Sayegh, K. (2018). Working and organizing in the age of the learning algorithm. Information and Organization, 28(1), 62-70

    work page 2018

  16. [16]

    S., & Pentland, B

    Feldman, M. S., & Pentland, B. T. (2003). Reconceptualizing organizational routines as a source of flexibility and change. Administrative Science Quarterly, 48(1), 94-118

    work page 2003

  17. [17]

    Galbraith, J. R. (1973). Designing Complex Organizations. Addison-Wesley

    work page 1973

  18. [18]

    Iyer, V., & Li, Z. (2023). LegalBench: A collaboratively built benchmark for measuring legal reasoning in large language models. Advances in Neural Information Processing Systems

    work page 2023

  19. [19]

    R., Lin, Z., Carroll, G

    Harrison, J. R., Lin, Z., Carroll, G. R., & Carley, K. M. (2007). Simulation modeling in organiza- tional and management research. Academy of Management Review, 32(4), 1229-1245

    work page 2007

  20. [20]

    Hecker, A. (2012). Knowledge beyond the individual? Making sense of a notion of collective knowl- edge in organization theory. Organization Studies, 33(3), 423-445

    work page 2012

  21. [21]

    Hong, L., & Page, S. E. (2004). Groups of diverse problem solvers can outperform groups of high- ability problem solvers. Proceedings of the National Academy of Sciences, 101(46), 16385-16389

    work page 2004

  22. [22]

    E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., & Narasimhan, K

    Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., & Narasimhan, K. (2024). SWE- bench: Can language models resolve real-world GitHub issues? International Conference on Learning Representations

    work page 2024

  23. [23]

    C., Valentine, M

    Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410

    work page 2020

  24. [24]

    W., Laubacher, R., & Dellarocas, C

    Malone, T. W., Laubacher, R., & Dellarocas, C. (2010). The collective intelligence genome. MIT Sloan Management Review, 51(3), 21-31. 18

    work page 2010

  25. [25]

    G., & Simon, H

    March, J. G., & Simon, H. A. (1958). Organizations. Wiley

    work page 1958

  26. [26]

    A., Veness, J., Bellemare, M

    Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Ried- miller, M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., & Hassabis, D. (2015). Human-level control through deep reinforcement learning. Nature, 518, 529-533

    work page 2015

  27. [27]

    Oborn, E., Barrett, M., & Dawson, S. (2013). Distributed leadership in policy formulation: A sociomaterial perspective. Organization Studies, 34(2), 253-276

    work page 2013

  28. [28]

    Orlikowski, W. J. (2000). Using technology and constituting structures: A practice lens for study- ing technology in organizations. Organization Science, 11(4), 404-428

    work page 2000

  29. [29]

    Orlikowski, W. J. (2007). Sociomaterial practices: Exploring technology at work. Organization Studies, 28(9), 1435-1448

    work page 2007

  30. [30]

    J., & Scott, S

    Orlikowski, W. J., & Scott, S. V. (2008). Sociomateriality: Challenging the separation of technol- ogy, work and organization. Academy of Management Annals, 2(1), 433-474

    work page 2008

  31. [31]

    T., Haerem, T., & Hillison, D

    Pentland, B. T., Haerem, T., & Hillison, D. (2010). Comparing organizational routines as recurrent patterns of action. Organization Studies, 31(7), 917-940

    work page 2010

  32. [32]

    Powell, W. W. (1990). Neither market nor hierarchy: Network forms of organization. Research in Organizational Behavior, 12, 295-336

    work page 1990

  33. [33]

    Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation- augmentation paradox. Academy of Management Review, 46(1), 192-210

    work page 2021

  34. [34]

    Ratner, H., & Plotnikof, M. (2022). Technology and dis/organization: Digital data infrastructures as partial connections. Organization Studies, 43(7), 1049-1067

    work page 2022

  35. [35]

    Riedl, C., De Cremer, D., Lucarelli, G., & Antoine-Souklaye, E. (2025). The potential and chal- lenges of AI for collective intelligence. Collective Intelligence, 4(1)

    work page 2025

  36. [36]

    R., Ben-Menahem, S

    Shrestha, Y. R., Ben-Menahem, S. M., & von Krogh, G. (2019). Organizational decision-making structures in the age of artificial intelligence. California Management Review, 61(4), 66-83

    work page 2019

  37. [37]

    Simon, H. A. (1947). Administrative Behavior. Macmillan

    work page 1947

  38. [38]

    L., & Griesemer, J

    Star, S. L., & Griesemer, J. R. (1989). Institutional ecology, translations, and boundary objects. 19 Social Studies of Science, 19(3), 387-420

    work page 1989

  39. [39]

    Stark, D., & Vanden Broeck, P. (2024). Principles of algorithmic management. Organization The- ory, 5(2)

    work page 2024

  40. [40]

    Steinberg, M. (2022). From automobile capitalism to platform capitalism: Toyotism as a prehistory of digital platforms. Organization Studies, 43(7), 1069-1090

    work page 2022

  41. [41]

    Symon, G., & Pritchard, K. (2015). Performing the responsive and committed employee through the sociomaterial mangle of connection. Organization Studies, 36(2), 241-263

    work page 2015

  42. [42]

    Thompson, J. D. (1967). Organizations in Action. McGraw-Hill

    work page 1967

  43. [43]

    Tsoukas, H. (1996). The firm as a distributed knowledge system: A constructionist approach. Strategic Management Journal, 17(S2), 11-25

    work page 1996

  44. [44]

    Tsvetkova, M., Yasseri, T., Pescetelli, N., & Werner, T. (2024). A new sociology of humans and machines. Nature Human Behaviour, 8, 1864-1876

    work page 2024

  45. [45]

    Weick, K. E. (1979). The Social Psychology of Organizing. Addison-Wesley

    work page 1979

  46. [46]

    Williamson, O. E. (1985). The Economic Institutions of Capitalism. Free Press

    work page 1985

  47. [47]

    W., Chabris, C

    Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., & Malone, T. W. (2010). Evidence for a collective intelligence factor in the performance of human groups. Science, 330(6004), 686-688. A Appendix A. Additional Simulation Evidence B Appendix B. Prompt Robustness and Prompt-Free Organiza- tion Selection Prompt robustness analysis addresses the possi...

    work page 2010