Summary
This paper presents a tool-augmented, MCP-based multi-agent system for job position normalisation, mapping unstructured job postings to standardised positions at scale.
- 95% accuracy on 14,000+ job positions
- 72.5% faster classification than baseline approaches
- 41.8% fewer API calls through shared caching
Architecture
The system orchestrates four specialised agents over the Model Context Protocol (MCP):
- Planner: decomposes the normalisation task
- Search: retrieves candidate standard positions
- Extraction: pulls structured attributes from postings
- Evaluation: scores and validates outputs
A shared caching layer across agents reduces redundant LLM calls while preserving accuracy.
Results
- Accuracy: 95%
- Speed: 72.5% faster than baselines
- Cost: 41.8% fewer API calls
References & Links
- ACM DL: https://dl.acm.org/doi/10.1145/3786335.3813207
- alphaxiv discussion: https://www.alphaxiv.org/abs/multi-agent-position-classification
- Author: Vahid Faraji, Applied AI Specialist
- Related paper: Tokalator: A Context Engineering Toolkit
- Related paper: Leveraging LLMs For Turkish Skill Extraction
Cite this work
@inproceedings{faraji2026multiagent,
title = {Multi-Agent Position Classification with Tool Orchestration:
Use Case System for Occupational Taxonomy Mapping},
author = {Faraji, Vahid},
booktitle = {Proceedings of the ACM Conference on AI and Agentic Systems (CAIS '26)},
year = {2026},
address = {San Jose, CA, USA},
publisher = {Association for Computing Machinery},
doi = {10.1145/3786335.3813207}
}