CONSTRAINT-AWARE MULTI-AGENT TASK SCHEDULING FOR AI-ENABLED ONE-PERSON COMPANIES IN DIGITAL CONTENT SERVICES

AI-enabled one-person companies use large language models and software agents to organize content planning, generation, review, and delivery with limited human resources. However, additional roles and model calls may raise cost and propagate errors without producing a stable quality gain. This paper presents a constraint-aware multi-agent task-scheduling prototype for digital content services and compares four execution configurations under controlled conditions. Six frozen tasks were processed by G1 one-shot generation, G2 same-role staged generation, G3 a fixed multi-agent workflow, and G4, which combined constraint-aware role selection, deterministic validator feedback, and bounded final-node revision. Each configuration was repeated three times per task, producing 72 formal runs and 406 API calls at a total cost of USD 0.221318. G4 achieved the highest automatic pass rate (72.22%) and validation score (0.9689), but also the highest mean call volume and cost. After Holm correction, the pass-rate difference remained significant only between G3 and G4; G4 validation scores exceeded G2 and G3 but not G1. Two non-author evaluators independently rated all 72 anonymized outputs. Averaged-score reliability was ICC(2,2)=0.7288, while human mean scores did not differ significantly across groups. Automatic validation was not significantly associated with human ratings. Because G4 changed several mechanisms together, the experiment does not identify their independent effects. The findings support a limited improvement in compliance with encoded rules, not a comprehensive human-perceived quality advantage.

Hali tarjima qilinmagan