Accountability Continuity in Persistent Multi-Agent AI Workforces: A Reference Architecture for Governed, Durable and Verifiable Delegated Work
This preprint presents a deployment-neutral reference architecture for preserving accountability across persistent multi-agent AI workforces. The paper addresses a central systems-engineering problem: how responsibility, bounded authority, execution state and verifiable evidence can remain connected as work is decomposed, delegated, transferred, executed, interrupted, resumed and accepted. The proposed architecture introduces a lifecycle-oriented design built around explicit work packages, validated task decomposition, bounded delegation, controlled execution, durable progress and recovery, evidence-based verification, conservative parent-level aggregation and auditable decision records. The manuscript identifies ten engineering requirements, defines the minimum contents of a transferable work package, describes the responsibilities and decision boundaries of the main architectural components, and analyses failure scenarios involving unauthorized delegation, stale approvals, partial execution, duplicate effects, missing evidence and premature completion. It also proposes an evaluation protocol covering ten scenario groups and measurable outcomes, including unauthorized effect rate, false completion rate, delegation integrity, evidence completeness, duplicate side-effect rate and recovery correctness. Research contribution: The paper provides a conceptual synthesis connecting established mechanisms across the full lifecycle of delegated multi-agent work. It does not claim that these individual mechanisms are new, nor does it claim empirical validation of the proposed architecture. Research methodology: Selective narrative literature review and conceptual systems-engineering analysis. Research status: Research preprint, Version 1.0. Not peer reviewed. No empirical results are reported. The architecture is deployment-neutral and is not tied to a particular software product, programming language, model provider or orchestration framework. Author: Chandra HothaORCID: 0009-0008-6353-0988
Authors
- Chandra Hotha
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24264516
- Primary Topic
- Access Control and Trust
- Type
- preprint