Multi-Agent Position Classification with Tool Orchestration: Use Case System for Occupational Taxonomy Mapping
Vahid Farajijobehdar (Kariyer.net R&D Center), İlknur Köseoğlu Sarı (Kariyer.net R&D Center), Nazım Kemal Üre (Stanford University), Engin Zeydan (Centre Tecnològic de Telecomunicacions de Catalunya)
Architectural Patterns & Composition
Summary
A confidence-gated multi-agent architecture using MCP that normalizes free-form job titles across five languages into occupational taxonomies, reducing classification time by 72.5%.
Description
Cross-lingual normalization of free-form job titles to standard occupational taxonomies is essential for labor market analytics, skill-gap detection, and regulatory compliance, but remains difficult because titles vary in language, script, and jargon. Existing approaches typically rely on single-pass pipelines that lack structured access to authoritative occupational databases and do not incorporate multi-agent orchestration or confidence-aware routing. To tackle this problem, this paper provides a confidence-gated multi-agent architecture built on the Model Context Protocol (MCP). This orchestrates specialized AI agents and external occupational databases to normalize noisy job titles across five languages into unified classification hierarchies. The architecture employs a hierarchical agent framework with a root Position Agent that delegates sub-agents for web search, position extraction, database queries, and trend analysis. The system is evaluated on 14K production positions and selected samples of 9,543 user-submitted free-text titles. The agent-assisted workflow reduces manual classification time from 4.0 min to 1.1 min per position (72.5% reduction, 3.6× speedup), with P50 latency of 1.2 s for API calls and 4.2 s for full agent orchestration with structured LLM output. The system achieves 95% segment accuracy with 0.94 average confidence.
