SCIENTRA — The Global Research Platform

Mediation, State, and Verification in Enterprise Multi-Agent AI Systems: A Formal Reference Architecture and Evaluation Protocol

Ahmed Abdullah

Published in SCIENTRA on 4 October 2026.

Abstract

Enterprise deployment of large language model agents is increasingly limited less by model capability than by the problem of coordinating models, tools, persistent business state, authorization policies, and human decisions into a workflow that can be operated and audited. Recent evidence complicates the assumption that multi-agent decomposition is beneficial by default: compute-normalized comparisons report that single agents match or exceed multi-agent systems at equal reasoning budgets, and failure analyses attribute a large share of multi-agent breakdowns to specification and inter-agent verification rather than to component capability. This paper takes that evidence as its starting point and asks a narrower question: which architectural properties, as opposed to which agent topology, determine whether an enterprise agentic workflow is governable, recoverable, and auditable. We formalize an enterprise agentic workflow as a policy-constrained stateful system over agents, tools, workflow state, authorization policies, and human decision points, and establish four elementary results. First, policy soundness is a property of complete mediation of external actions and is independent of agent policy quality, which implies that it survives prompt injection and model failure whereas prompt-level guardrails do not. Second, recoverability and auditability require a durable external projection of every state component on which a consequential decision depends; context-resident state admits observationally indistinguishable histories with divergent correct continuations. Third, under an explicit independence idealization, sequential decomposition into n bounded agents without inter-stage verification bounds end-to-end success by pn, so the value of decomposition is conditioned on handoff verification rather than on specialization, which is consistent with reported empirical failure taxonomies. Fourth, attributability follows from complete mediation together with a total authorization relation over consequential actions and append-only logging. We map these results onto a five-layer reference architecture, a six-stage adoption protocol with explicit exit evidence, and a multi-level evaluation framework. We then specify a comparative evaluation protocol and an accompanying task schema that make the architecture falsifiable, including the configurations, metrics, and reporting requirements needed to test whether coordination earns its cost. The analytic results are proved; the architectural and adoption claims are stated as testable propositions and are not empirically validated here. We state explicitly which claims require experiment and what evidence would refute them.

Keywords

multi-agent systemsLLM agentsreference architectureAI governanceagent evaluationcomplete mediationenterprise workflow automation

References

  1. Autio, C., Schwartz, R., Dunietz, J., Jain, S., Stanley, M., Tabassi, E., Hall, P., & Roberts, K. (2024). Artificial
  2. Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1.
  3. National Institute of Standards and Technology. https://doi.org/10.6028/NIST.AI.600-1
  4. Barres, V., Dong, H., Ray, S., Si, X., & Narasimhan, K. (2025). τ 2-Bench: Evaluating Conversational Agents
  5. in a Dual-Control Environment. arXiv:2506.07982.
  6. Cemri, M., Pan, M. Z., Yang, S., Agrawal, L. A., Chopra, B., Tiwari, R., Keutzer, K., Parameswaran, A., Klein, D., Ramchandran, K., Zaharia, M., Gonzalez, J. E., & Stoica, I. (2025). Why Do Multi-Agent LLM Systems Fail? Proceedings of the 42nd International Conference on Machine Learning (ICML). arXiv:2503.13657.
  7. Chan, A., Ezell, C., Kaufmann, M., Wei, K., Hammond, L., Bradley, H., Bluemke, E., Rajkumar, N., Krueger, D., Kolt, N., Heim, L., & Anderljung, M. (2024). Visibility into AI Agents. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 958–973. https: //doi.org/10.1145/3630106.3658948
  8. Chan, A., Wei, K., Huang, S., Rajkumar, N., Perrier, E., Lazar, S., Hadfield, G. K., & Anderljung, M. (2025).
  9. Infrastructure for AI Agents. arXiv:2501.10114.
  10. European Commission (2025). Guidelines on the scope of obligations for providers of general-purpose AI
  11. models; General-Purpose AI Code of Practice; Template for the public summary of training content. Brussels.
  12. European Parliament and Council of the European Union (2024). Regulation (EU) 2024/1689 of 13 June
  13. 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal
  14. of the European Union, L, 12 July 2024.
  15. Huang, K.-H., Prabhakar, A., et al. (2025). CRMArena-Pro: Holistic Assessment of LLM Agents Across
  16. Diverse Business Scenarios and Interactions. arXiv:2505.18878.
  17. Kolt, N. (2025). Governing AI Agents. Notre Dame Law Review, 101 (forthcoming). arXiv:2501.07913.
  18. Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., et al. (2024). AgentBench: Evaluating LLMs as Agents. International Conference on Learning Representations (ICLR). arXiv:2308.03688.
  19. Lu, J., Holleis, T., Zhang, Y., Aumayer, B., Nan, F., Bai, F., Ma, S., et al. (2024). ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities. arXiv:2408.04682.
  20. National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework
  21. (AI RMF 1.0). NIST AI 100-1. https://doi.org/10.6028/NIST.AI.100-1
  22. Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative Agents:
  23. Interactive Simulacra of Human Behavior. Proceedings of the 36th Annual ACM Symposium on User
  24. Interface Software and Technology (UIST). arXiv:2304.03442.
  25. Plaat, A., van Duijn, M., van Stein, N., Preuss, M., van der Putten, P., & Batenburg, K. J. (2025). Agentic
  26. Large Language Models: A Survey. arXiv:2503.23037.
  27. Saltzer, J. H., & Schroeder, M. D. (1975). The Protection of Information in Computer Systems. Proceedings
  28. of the IEEE, 63(9), 1278–1308.
  29. Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., & Scialom, T. (2023). Toolformer: Language Models Can Teach Themselves to Use Tools. Advances in Neural Information Processing Systems (NeurIPS). arXiv:2302.04761.
  30. Shu, R., Das, N., Yuan, M., Sunkara, M., & Zhang, Y. (2024). Towards Effective GenAI Multi-Agent Collaboration: Design and Evaluation for Enterprise Applications. arXiv:2412.05449. (Industry technical
  31. report; not peer reviewed.)
  32. Staufer, L., Feng, K. J. K., Wei, K., Bailey, L., Duan, Y., Yang, M., Ozisik, A. P., Casper, S., & Kolt, N. (2026). The 2025 AI Agent Index: Documenting Technical and Safety Features of Deployed Agentic AI Systems. ACM Conference on Fairness, Accountability, and Transparency (FAccT). https://doi. org/10.1145/3805689.3806728
  33. The White House (2025). Winning the Race: America’s AI Action Plan. Executive Office of the President,
  34. Washington, DC, 23 July 2025.
  35. Tran, D., & Kiela, D. (2026). Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning
  36. Under Equal Thinking Token Budgets. arXiv:2604.02460.
  37. Tran, K.-T., Dao, D., Nguyen, M.-D., Pham, Q.-V., O’Sullivan, B., & Nguyen, H. D. (2025). Multi-Agent Collaboration Mechanisms: A Survey of LLMs. arXiv:2501.06322.
  38. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain of
  39. Thought Prompting Elicits Reasoning in Large Language Models. Advances in Neural Information Processing Systems (NeurIPS). arXiv:2201.11903.
  40. Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., Awadallah, A. H., White, R. W., Burger, D., & Wang, C. (2024). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. Conference on Language Modeling (COLM). arXiv:2308.08155.
  41. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. International Conference on Learning Representations (ICLR). arXiv:2210.03629.
  42. Yao, S., Shinn, N., Razavi, P., & Narasimhan, K. (2024). τ -bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. arXiv:2406.12045.
  43. Zhang, et al. (2025). Single-Agent or Multi-Agent Systems? Why Not Both? arXiv:2505.18286.
  44. Zhu, Y., et al. (2025). Establishing Best Practices for Building Rigorous Agentic Benchmarks. arXiv:2507.02825.