BU68 AAI-01|Agent Production Standing Architecture-Level Closure Criteria for Reliable Tool-Using LLM Agents

Read full text
Authors: Dedong Zhan
Cited by: 4
References: 3

Abstract

This paper develops BU68 AAI-01|Agent Production Standing as the first file in the B_U-based Agentic AI series. Its central claim is that the main bottleneck of Agentic AI is not the absence of more autonomous or more fluent models, but the absence of production-standing criteria. A demo can succeed inside a narrow projection closure while still failing in production because production requires state continuity, tool closure, residual recovery, cost boundedness, verification, auditability, and reproducible settlement. The paper reclassifies common Agent failures as structural closure failures. Infinite loops are treated as unsettled residuals returning into the execution cycle. State collapse is redefined as same-task-world drift. Unstable tool use is identified as unverified observation writeback. High inference cost is traced to misrouted information-update cost. Reproducibility failure is treated as the absence of a settlement ledger rather than a mere logging problem. BU68 introduces Agent Production Standing as a runtime admissibility condition. A production-standing Agent must preserve the same task world during updates, close every tool call through observation and verification, relocate errors into explicit recovery actions, remain inside cost boundaries, preserve traceability, and deliver results that can be replayed and checked. The paper further distinguishes rules from structure: rules can reduce residual exposure, but architecture-level closure reduces residual generation. Excessive rule accumulation is therefore treated as a symptom of structural incompleteness. The final judgment is that reliable Agentic AI requires a shift from prompt rules, autonomy levels, and demo success toward architecture-level settlement. Production standing is delivered by task-runtime alignment, state architecture, tool closure, residual routing, cost gates, verification chains, and settlement ledgers. A model may provide local readout; production standing must be achieved by a closed, auditable, and reproducible runtime structure.

Fields of Study & Subjects

Multi-Agent Systems and Negotiation AI-based Problem Solving and Planning Mobile Agent-Based Network Management

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
9 May 2026
Document type
Preprint