{
  "researchId": "JPC-26-00142",
  "researchUrl": "https://jipcet.org/r/JPC-26-00142",
  "articleUrl": "https://jipcet.org/articles/jipcet-2026-0142",
  "articleRef": "JIPCET-2026-0142",
  "externalDoi": null,
  "title": "Evaluating Agentic AI Architectures for Enterprise Applications: A Multi-Site Controlled Comparison of Orchestration Patterns",
  "authors": [
    {
      "name": "Ruth Okonkwo",
      "affiliation": "University of Wollongong, Australia",
      "orcid": null
    },
    {
      "name": "Hannah Lindqvist",
      "affiliation": "University of Missouri–Kansas City, USA",
      "orcid": null
    },
    {
      "name": "Simon Marsh",
      "affiliation": "Universiti Teknologi Malaysia, Malaysia",
      "orcid": null
    },
    {
      "name": "Elena Vasquez",
      "affiliation": "University of Missouri–Kansas City, USA",
      "orcid": null
    }
  ],
  "abstract": "Background. Enterprise deployment of large language model agents has advanced faster than the evidence base needed to choose between orchestration architectures, and most published comparisons rely on single-task benchmarks that do not resemble production workflows. Objective. We quantify how three orchestration patterns — unconstrained sequential chaining, supervisor–worker delegation, and constrained tool-calling with schema-validated actions — differ in task completion, latency, cost per completed task, and failure composition. Methods. We conducted a pre-registered, paired-design evaluation across twelve enterprise workflows drawn from finance operations, IT service management, procurement, and clinical scheduling at four organisations, yielding 4,800 task instances. Each instance was executed under all three architectures with identical model checkpoints, temperature, retrieval corpora and tool inventories. Completion was scored by two independent domain raters (Cohen's κ = 0.83) with adjudication by a third. We repeated the full protocol on three model checkpoints to test ordinal stability, and fitted mixed-effects logistic models with random intercepts for workflow and organisation. Results. Constrained tool-calling completed 0.81 of tasks (95% CI 0.79–0.83) against 0.74 for supervisor–worker delegation and 0.62 for unconstrained chaining, a 31% relative improvement over chaining. Prompt-parity controls recovered only 3.1 of the 19 percentage-point gap, indicating that the advantage is architectural rather than a prompt-engineering artefact. Constrained runs cost 4% more per model call but 17% less per completed task. The architectural ordering held on all three checkpoints while absolute completion moved by up to 11 points. Conclusion. Constraint at the action boundary, not model capability alone, explains most reliability variance in enterprise agent deployments; procurement and platform decisions should be evaluated per completed task rather than per call.",
  "keywords": [
    "AI agents",
    "large language models",
    "enterprise AI",
    "evaluation methodology",
    "orchestration architecture",
    "tool calling"
  ],
  "area": "AI & Intelligent Systems",
  "issue": "JIPCET 2026; 3(2)",
  "pageRange": "1–40",
  "published": "2026-07-22",
  "received": "Received 04 Feb 2026",
  "version": "1.0",
  "versions": [
    {
      "version": "1.0",
      "date": "2026-07-22",
      "note": "Version of record published after editorial decision."
    }
  ],
  "articleType": "Research Article",
  "reviewStatus": "Peer reviewed under named human editorial supervision",
  "peerReviewed": true,
  "handlingRecord": "Handled by a named JIPCET editor of record, who read the manuscript, oversaw review and approved the decision.",
  "methodology": "Research Article",
  "dataAvailability": "Available from the corresponding author on reasonable request, in line with participant and organisational agreements.",
  "codeAvailability": "Analysis procedures are documented in full in the methods section of the article.",
  "referenceCount": 8,
  "references": [
    "Okonkwo, R., Marsh, S. & Paxton, D. (2025). Reliability bounds for compositional LLM agents. Journal of Emerging Computing 18(1), 44–61.",
    "Marsh, S. & Hale, E. (2024). Tool-calling constraints and malformed action rates in production agents. Proceedings of APSEC, 202–215.",
    "Rahman, A. (2026). Evaluation cost models for multi-step inference workloads. JIPCET 1(1), 12–29.",
    "Vasquez, E. & Paxton, D. (2025). Paired designs for agent benchmarking. Empirical Methods in Applied AI 4(2), 133–158.",
    "Tan, M. & Haddad, J. (2025). Per-task accounting for serverless model invocation. Proceedings of SoCC, 288–301.",
    "Whitmore, O. (2025). Reliance, oversight and calibration in automated workflows. Human Factors 67(2), 210–229.",
    "Russo, G. & Whitfield, B. (2024). Quality proxies and the unit-of-measurement problem. Empirical Software Engineering 29(1), 77–98.",
    "Ferreira, L. (2026). Where recovered time goes: validation load after automation. JIPCET 2(3), 40–63."
  ],
  "replications": "No independent replication recorded to date.",
  "corrections": "No corrections issued. Any correction is published as a new version with a dated entry in the version history.",
  "license": "CC BY 4.0",
  "fullTextHtml": "https://jipcet.org/articles/jipcet-2026-0142",
  "fullTextPdf": "https://jipcet.org/papers/jipcet-2026-0142.pdf",
  "publisher": {
    "name": "JIPCET Editorial Office",
    "url": "https://jipcet.org",
    "email": "editor@jipcet.org"
  },
  "journal": {
    "title": "Journal of Innovation, Product, Computing & Emerging Technologies",
    "short": "JIPCET"
  }
}