2026.10.31 Continuous Operation in the GL Mac Environment: Current State of Local LLMs and Subscription AI

Setup and Workload Allocation across 11 Macs

The GL Mac environment operates across 11 Macs in the office: 1 coordinator and 10 workers. Our design intent clearly separates roles: local LLMs handle standard drafting, while contracted subscription AI (Claude Code, Codex, and Grok) handles judgment and final polish. Subscription AI is used strictly within contracted subscription limits, additional usage is disabled, and no API billing paths are created. To prevent operations from halting at subscription limits, the design has Clef (a local judging model) and local LLMs fill the gaps.

Operational Preparations for Continuous Processing

In the GL Mac environment, we combine local LLMs with contracted subscription AI to prepare the operational foundation for continuous marketing and development tasks. The goal of this initiative is to make services available 24 hours a day. However, constant 100% utilization of GPUs or paid AI is not our objective. We prioritize placing resources on standby during idle periods to ensure steady processing when needed.

Definition of Operational Status and Current Pipeline Suspension

As of October 26, 2026, both 24-hour generation pipelines remain offline, though conditions may change by the time this is published. We define an operational state strictly when two conditions are met: active production callers exist, and a new row of output appears within the scheduled period. Mere background process execution is not considered an operational state. The causes differ by pipeline. One pipeline halted because quality-gate rejections were counted as consecutive failures. The other halted because the denominator of the work-in-progress (WIP) limit was a running total that approvals never reduced. The recovery steps also differ: the former requires a daily cap on unattended finishing, after which a human resets the failure counter; the latter requires correcting the denominator so that it counts only pending drafts. At the time of writing, neither pipeline had been recovered. The immediate next step is the recovery of the halted generation pipelines.

Publication is Approved and Executed by Humans

In the GL Mac environment, publication is approved and executed by humans. AI does not publish deliverables autonomously. Even when generation occurs automatically under 24-hour operation, final verification and publishing operations are carried out by humans. We maintain a design where publishing authority and responsibility strictly remain with humans.

Note: The text and image of this article were created with AI assistance; humans reviewed and confirmed them before publication.