Supplementary materials for "Organizationally-Intelligent Agents: Connecting Multi-Agent Engineering Systems to Organizational Design Theory"
This repository contains the corpus list, attribute classifications, and taxonomy rationale for the 30 multi-agent system (MAS) frameworks analyzed in the paper. These materials support reproducibility by enabling other researchers to replicate our data collection and classification methodology.
| File | Description |
|---|---|
corpus_summary.csv |
Citation keys, titles, one-line descriptions, domains, and taxonomy classifications for all 30 frameworks |
full_attributes.csv |
Complete 8-attribute classification for each framework (orchestration, organizational structure, reasoning protocols, operational methodologies, memory, modality, agent count, Asimow phase coverage) |
classification_rationale.md |
Taxonomy definitions and per-paper rationale explaining classification decisions with key identifying features |
coding_protocol.md |
Detailed coding protocol describing how each attribute was extracted and how taxonomy classifications were assigned, including handling of omitted details |
od_mapping_methodology.md |
Methodology for mapping taxonomy categories to organizational design concepts, including selection criteria and validation approach |
- Orchestrated (16, 53%): Explicit orchestrator/manager agent coordinates other agents
- Networked (8, 27%): Peer agents communicate via shared state or message passing without central coordinator
- Sequential (6, 20%): Pipeline/workflow with predetermined execution order and stage-to-stage handoffs
- Iterative-Feedback (15, 50%): Repeated generation-evaluation-refinement cycles until convergence
- Graph-Routed (8, 27%): Dynamic execution paths determined by routing logic or state machines
- Staged-Pipeline (7, 23%): Fixed sequence of distinct phases with prescribed transitions
| Sequential | Networked | Orchestrated | |
|---|---|---|---|
| Iterative-Feedback | 1 | 6 | 8 |
| Graph-Routed | 1 | 2 | 5 |
| Staged-Pipeline | 4 | 0 | 3 |
Inclusion criteria: Multi-agent systems employing LLMs or AI agents for hardware/physical product design (mechanical, structural, circuit, fluid mechanics, materials discovery) with sufficient description of agent architecture, task decomposition, or organizational structure.
Exclusion criteria: Software engineering (mature tooling with distinct patterns), robotic RL agents (distinct action spaces/reward structures), pure control theory/optimization without design synthesis.
Search strategy: Three keyword sets ("Multi-Agent Framework for engineering design," "AI Agents in Engineering Design," "Agentic Frameworks for Engineering Design") across arXiv, IEEE Xplore, ACM Digital Library, Design Society, and ASME IDETC-CIE proceedings. Searches targeted 2024-2025 publications and were conducted November-December 2024.
If you use these materials, please cite:
@inproceedings{ezemba2025organizationally,
title={Organizationally-Intelligent Agents: Connecting Multi-Agent Engineering Systems to Organizational Design Theory},
author={Ezemba, Jessica and McComb, Christopher and Tucker, Conrad},
article={Design Computing and Cognition'26},
year={2025}
}For questions about the classification methodology or to report issues, please contact the corresponding author: jezemba@andrew.cmu.edu