ryanyen22/multi-agent-systematic-review

Systematic Literature Review: Multi-Agent Collaboration — Frameworks, Libraries, Languages, and Protocols

A reflexive thematic analysis of 107 papers/frameworks/protocols identifying shared primitives and interaction patterns across the multi-agent AI ecosystem (2022–2025)

Overview

This repository contains a systematic literature review conducted using reflexive thematic analysis (Braun & Clarke, 2006) with constant comparison methodology. The analysis moved iteratively between the dataset, coded extracts, and emerging themes across three documented iterations, each with full rationale for changes.

Key Contribution

After analyzing 107 works, we identify that every multi-agent system in the corpus reduces to combinations of the same fundamental primitives — different papers merely:

1. Use different terminology for the same underlying mechanism
2. Make different primitives "first-class" vs implicit
3. Compose them via different meta-mechanisms (code, prompts, protocols, latent states)

Repository Structure

├── README.md                          # This file
├── corpus.md                          # Full 107-paper corpus with metadata + interaction tags
├── iteration_1_initial_coding.md      # First pass: 34 raw codes
├── iteration_2_theme_generation.md    # Second pass: collapse synonyms, 24 themes emerge
├── iteration_3_final_refinement.md    # Third pass: multi-level hierarchy, 15 primitives survive
├── version_changelog.md              # What changed between iterations and WHY
├── what_they_actually_do.md          # Cutting through buzzwords: what each framework really does
├── interaction_taxonomy.md           # Full interaction pattern classification
└── synthesis.md                      # Final integrated synthesis

Interaction Pattern Tags

Tag	Pattern	Definition
human-AI	1 human ↔ 1 AI	Single user working with a single AI agent/copilot
human*-AI	N humans → 1 AI	Multiple stakeholders interacting with shared AI system
human-AI*	1 human → N AI	Single user orchestrating a team of AI agents
AI-AI	AI ↔ AI	Agents communicating autonomously without human in active loop

How to Read This

1. Start with what_they_actually_do.md if you want to cut through marketing language
2. Read version_changelog.md to understand how themes evolved
3. Read synthesis.md for the final integrated findings
4. Use corpus.md as reference — every claim is traceable to specific papers

License

CC-BY-4.0