Where Agent Protocols Stop: Cross-Layer Gaps in Authority, Commitment, Evidence, and Service Closure
- 1Harbin Institute of Technology
Abstract
Local protocol success is not sufficient to close an agent service. We study when locally conformant protocol states justify relationship-level service closure across agent stacks. A protocol-neutral Agent Service Episode (ASE) separates capability, authority, commitment, process, authoritative external effect, evidence/acceptance, responsibility, and relationship state. The protocol corpus uses a specification snapshot cutoff of 15 September 2026. Two authors independently coded nine semantic-ownership dimensions. On 32 held-out cells, exact agreement was 78.1%, Cohen's κ=0.566 (bootstrap 95% CI $[0.295,0.815]$), and nominal Krippendorff's α=0.564; the held-out sample contained no Delegated labels, so D/U reproducibility is not estimated. All six families natively own action/process semantics, but none owns or names an external owner for authoritative external effect or service closure at ASE scope; all 12 corresponding raw judgments were independently U/U. An Alloy upper-bound model shows that the strongest observed local semantics do not entail ASE closure. An integrated binding model establishes bounded joint realizability and single-removal non-redundancy, with D0 used as an encoding-alignment regression. An executable Agent Service Binding Profile demonstrates the same bindings on invariant-targeted CAD/CAE/PLM traces in a coverage/feasibility study. Thus, within this protocol snapshot, TaskCompleted does not by itself license Closed. For Agentic Services Computing, service closure is a first-class cross-layer governance decision whose owner and validation rule must be explicit.
Related
- AI-driven HR service quality and organisational citizenship behaviour in the Indian IT sector: the mediating role of organisational commitment
- Detector-Facing Empirical Protocols for Semantic Commitment Quantum Gravity
- Legal Evidence–Authority Dissociation in juror judgment: a multistudy reanalysis of memory contamination and AI advice
- Transcriptomic Reproducibility in Endometrial Biopsies: Evidence Across Cycles and Tissue Regions in Hormonal Replacement Therapy Endometrial Preparation Protocols
- The Weight of the Past: Sunk Costs, Escalation of Commitment, and the Limits of the Evidence
BibTeX
@misc{wang2026where,
title = {{Where Agent Protocols Stop: Cross-Layer Gaps in Authority, Commitment, Evidence, and Service Closure}},
author = {Zhongjie Wang and Mingyi Liu and Shuang Yu and Xiang He},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23448113},
url = {https://zenodo.org/records/23448113},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Where Agent Protocols Stop: Cross-Layer Gaps in Authority, Commitment, Evidence, and Service Closure
AU - Zhongjie Wang
AU - Mingyi Liu
AU - Shuang Yu
AU - Xiang He
PY - 2026
DA - 2026/10/18
PB - Zenodo
DO - 10.5281/zenodo.23448113
UR - https://zenodo.org/records/23448113
AB - Local protocol success is not sufficient to close an agent service. We study when locally conformant protocol states justify relationship-level service closure across agent stacks. A protocol-neutral Agent Service Episode (ASE) separates capability, authority, commitment, process, authoritative external effect, evidence/acceptance, responsibility, and relationship state. The protocol corpus uses a specification snapshot cutoff of 15 September 2026. Two authors independently coded nine semantic-ownership dimensions. On 32 held-out cells, exact agreement was 78.1%, Cohen's κ=0.566 (bootstrap 95% CI $[0.295,0.815]$), and nominal Krippendorff's α=0.564; the held-out sample contained no Delegated labels, so D/U reproducibility is not estimated. All six families natively own action/process semantics, but none owns or names an external owner for authoritative external effect or service closure at ASE scope; all 12 corresponding raw judgments were independently U/U. An Alloy upper-bound model shows that the strongest observed local semantics do not entail ASE closure. An integrated binding model establishes bounded joint realizability and single-removal non-redundancy, with D0 used as an encoding-alignment regression. An executable Agent Service Binding Profile demonstrates the same bindings on invariant-targeted CAD/CAE/PLM traces in a coverage/feasibility study. Thus, within this protocol snapshot, TaskCompleted does not by itself license Closed. For Agentic Services Computing, service closure is a first-class cross-layer governance decision whose owner and validation rule must be explicit.
ER -