> ## Content Index > Fetch the complete content index at: https://thomasadair.ghost.io/llms.txt > Use this file to discover other available public pages before exploring further. # The Folder Is the Orchestrator - URL: https://thomasadair.ghost.io/folder-is-the-orchestrator/ - Published: 2026-09-09T18:36:40.000Z - Updated: 2026-09-09T18:36:40.000Z - Description: The folder is the orchestrator until concurrency earns the framework. - Author: Thomas Arthur Adair - Tags: short-form-note, #short-form-note The folder is the orchestrator until concurrency earns the framework. Most of what people reach for a multi-agent framework to do, a directory tree already does — numbered folders for stages, a markdown file at each stage telling one agent what job it’s doing, plain scripts for the mechanical parts. No orchestration layer. No message bus. The structure *is* the coordination. There’s a paper for it. Van Clief and McDermott published *Interpretable Context Methodology* back in March, naming the pattern Model Workspace Protocol: replace framework-level orchestration code with filesystem structure. Folders are pipeline stages. Markdown carries the role and the context. Local scripts do the non-AI work. They draw the lineage from Unix pipes and multi-pass compilers, which is the right lineage — this is an old idea wearing current clothes. I came across it after the fact, and two things struck me. First, it described how I’d already been working — the same shape, arrived at from a different direction. Second, it has an edge the popular version tends to skip: the folder holds only up to the point the work runs concurrently. Past there, the idea has to be paired with something that can actually coordinate. My version is three layers. A map file at the repo root that says what this is and where to look. A live-state file that holds what’s actually happening right now — the current job, what’s blocked, what got decided. Then the work folders. That’s the same shape the method describes — an index, a live-context file, the material — and its context file is close to a one-for-one match with the continuation brief I keep per project. ## The part that makes it more than a preference Structure isn’t just tidier. It retrieves better, and somebody measured it. STAIR, out of IBM this September, treats a document’s table of contents as the retrieval index — feed the whole ToC in with the query and constrain the model to answer with a real entry from it. Against flat chunking, where you slice a corpus into fixed-length pieces and match on similarity, it isn’t close. Across eighteen books in six domains, Recall@1 lands at 82.6% — against 76.9 for a differentiable search index, 68.7 for dense passage retrieval, 59.5 for BM25\. The bigger gap is elsewhere: hallucination drops to 0.05% from that search index’s 3.25%. That second number is the one I care about. It’s near-zero for the same reason my memory bundle is: if the model can only answer with a node that exists, it can’t invent one. Cite, don’t recall. Length-based chunks lose because a chunk boundary is arbitrary. A section title isn’t. Structure carries meaning, and the meaning survives retrieval — so ICM asserts folder-as-architecture, and STAIR measures why it works. ## Where the paper stops The paper is scoped, and almost everyone repeating it drops the scope. Van Clief and McDermott target *sequential* workflows where a human reviews each step. The abstract explicitly grants that multi-agent frameworks work well for complex, concurrent systems. That’s in the source. It’s not a hedge someone added later. The popular version has sanded that off. The methodology got picked up and turned into a starter kit and a free-course funnel, and somewhere in that compression, don’t-pay-framework-overhead-where-a-folder-does-the-job became a claim the paper doesn’t make: that multi-agent frameworks are a con. Those are different claims. The first is true and useful. The second contradicts the paper it cites. My own setup is the counterexample to that “frameworks are a con” claim. For most of what I do, the folder is more than enough. But the moment two processes run at once and have to agree on the same state, a folder has nothing to coordinate with — a directory listing can’t arbitrate. That is exactly where I’m building a real framework: a multi-agent bus with a shared canon, a queue, and an event log. Not because a framework looks serious — because concurrency is a harder problem than a folder was ever built to solve. Same shape, different physics: sequential work I read a step at a time; concurrent work needs a referee, and a folder isn’t one. ## What to take Try the folder first. It costs a directory and a markdown file, it’s legible to you six weeks later, and you can read the whole system without running it. If it holds, you never needed the framework. Add the framework when concurrency actually shows up — when two things run at the same time and have to agree about state. That’s the line worth watching for, and it arrives later than most people think. None of this is a trick of the tooling. The folder works because the work is legible — because someone can look at the structure and know what to do without being told. That held when the someone was a person handed a folder and left to get on with it. It holds now that it’s a model reading its way through the same tree. Build the system so the structure carries the instruction, and most of the time you’ll find you never reached for the framework at all. ## Sources - Jake Van Clief & David McDermott, *Interpretable Context Methodology: Folder Structure as Agentic Architecture*, arXiv:2603.16021 (submitted 17 Mar 2026; v2 18 Mar 2026; MIT license, CC-BY) — https://arxiv.org/abs/2603.16021 - Kumar, Pulivarthi, Kumar, Sen, Bhat & Joshi, *STAIR (STructure Aware Information Retriever): A novel dataset and LLM based retriever for document structure augmentation*, arXiv:2609.03874 (IBM; submitted 3 Sep 2026) — https://arxiv.org/abs/2609.03874 — SearchTome benchmark, 18 books across 6 domains: Recall@1 82.6% (DSI 76.9 / DPR 68.7 / BM25 59.5); non-leaf-node / hallucination rate 0.05% vs DSI 3.25%. - Ryan McKinney, filesnfolders.com — ICM starter kit and the Index / Context / Material mnemonic (popularization).