# Interpretable Context Methodology: Folder Structure as Agentic Architecture

**Authors:** Jake Van Clief, David McDermott  
**arXiv:** [2603.16021](https://arxiv.org/abs/2603.16021)  
**Date:** 2026-03-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2603.16021v1/

## Key points

- Traditional AI agent orchestration methods involve building complex multi-agent frameworks
- Model Workspace Protocol (MWP) proposed by the authors uses filesystem structure for orchestration
- Numbered folders represent different stages of the workflow in MWP
- Plain markdown files contain prompts and context cues for AI agent tasks in MWP
- Local scripts handle mechanical tasks that do not require AI intervention
- MWP enables a single agent to perform tasks without a complex multi-agent framework
- Draws inspiration from Unix pipeline design, modular decomposition, multi-pass compilation, and literate programming principles
- Protocol is open source under the MIT license for widespread adoption and customization
- Simplifies AI agent orchestration and enhances human-AI collaboration by reducing complexity and streamlining workflow automation

## Summary

The paper "Interpretable Context Methodology: Folder Structure as Agent Architecture" by Jake Van Clief and David McDermott from Eduba, University of Edinburgh, USA, presents a novel approach to AI agent orchestration. Traditional methods involve building complex multi-agent frameworks to manage context passing, memory, error handling, and step coordination through code. However, these frameworks introduce unnecessary engineering overhead in sequential workflows where human intervention is required at each step. To address this issue, the authors propose the Model Workspace Protocol (MWP) which utilizes filesystem structure instead of framework-level management to streamline orchestration. In this protocol, numbered folders represent different stages of the workflow while plain markdown files contain prompts and context cues for a single AI agent to execute specific tasks at each step. Additionally, local scripts handle mechanical tasks that do not require AI intervention. By adopting MWP, the system enables a single agent to perform tasks by reading designated files at the appropriate moment without the need for a complex multi-agent framework. This methodology draws inspiration from Unix pipeline design, modular decomposition, multi-pass compilation and literate programming principles to effectively structure context for AI agents in a more efficient manner. Furthermore, the protocol is open source under the MIT license allowing for widespread adoption and customization. This innovative approach not only simplifies AI agent orchestration but also enhances human-AI collaboration by reducing unnecessary complexity and streamlining workflow automation processes.