The ICM Paper, in plain English.

Ryan “RyMac” McKinney · USMC · 21 August 2026 · 4 minute read

The ICM paper is called Interpretable Context Methodology: Folder Structure as Agentic Architecture. Jake Van Clief and David McDermott put it up on 17 March 2026. I am not a researcher, I cannot write code, and I run the thing it describes every day. So here is what it says, without the language that makes people close the tab.

The one sentence version

Most people coordinate AI agents by writing code. the paper argues that for a lot of real work, the folder structure can do that job instead, and you can read a folder. That readable part is where the word interpretable comes from.

What it is careful to say, and what everybody misquotes

The paper does not say multi agent frameworks are bad. It says they work well for complex, concurrent systems.

The argument is narrower and stronger than that. For sequential work where a human looks at the output at each step, those frameworks add engineering the problem never asked for. That is most of what a business actually does, and it is the distinction I keep coming back to in agentic context engineering.

The name it gives its own method

The paper calls the method the Model Workspace Protocol, or MWP. You may hear people say protocol and assume software has to be installed. It simply means an agreed way of laying things out, the way a filing standard is a protocol.

3 parts, and I run all 3, and they map onto the folders I actually keep:

1. Numbered folders stand for stages. Not categories. Stages, in order. This is the single thing I did not understand for a month, and it is why my own folders read 01, 02, 03 rather than sitting in a neat alphabet.

2. Plain markdown files carry the instructions and the context. Markdown is a text file where a hash mark makes a heading. That is the entire technology. Each file tells one agent what part it is playing at that stage.

3. Ordinary scripts do the mechanical work. Renaming, moving, counting. The paper is direct that this work does not need AI at all, and paying a model to do it is waste. That line alone changed what I stopped paying for, which I broke down in Claude free vs pro.

Why numbering the folders matters more than it looks

A category tells the agent what a thing is. A stage tells it where in the job it is standing. Those are different, and the second one is what stops the guessing.

When my folders were named by category, the agent would helpfully start at whatever looked most interesting. Numbered, it walks. The difference in how much desk space that saves is not small.

If you would rather see that than read it, I let the folders audit themselves on camera with no script, and it found 2 of my own files contradicting each other.

Where the ideas come from

The paper is open that it borrowed. It names Unix pipeline design, modular decomposition, multi-pass compilation, and literate programming.

Those are 4 phrases that will lose most readers, so here they are as one idea: do one thing, hand the result to the next thing, and write it so a human can read it. Unix has worked that way since the 1970s. None of this is new, and the paper does not pretend it is. That honesty is a large part of why I trust it, and it is part of what I said in is Jake Van Clief legit.

It is free

The protocol is open source under the MIT license, which means you can use it, change it, and build a business on it without asking anybody. Read it yourself. It is not long, and skipping the parts that lose you is allowed.

Jake teaches the folder version at Clief Notes, and that is a paid room and that is an affiliate link. The method underneath it costs nothing and always did.

What I would tell you to do with it

Read the abstract. Skip anything that loses you. Then go set up 2 text files and 4 folders and run 1 real job through them, because you will learn more in that hour than in a second read.

If you want mine rather than a blank one, the template is here. It is the paper's shape with my work already in it.