# Interpretable Context Methodology (ICM)

Using folder structure — plain files in a well-organized hierarchy — as the orchestration mechanism for a multi-step AI workflow, instead of a code-level framework.

**Interpretable Context Methodology (ICM)**, also called *folder structure as agent architecture*, is the discipline of using filesystem layout — folders, plain Markdown files, and stage-scoped context contracts — as the orchestration mechanism for a multi-step LLM workflow, in place of a code-level multi-agent framework. It was introduced by [Jake Van Clief](/wiki/jake-van-clief.md) and [David McDermott](/wiki/david-mcdermott.md) in [their 2026 paper](/docs/icm-paper.md).

## The core idea

If the prompts and context for each stage of a workflow already exist as files in a well-organized folder hierarchy, you do not need multiple agents or a coordination framework — you need one agent that reads the right files at the right moment. The folder structure *is* the orchestration. The argument borrows directly from Unix philosophy: programs that do one thing, the output of one becoming the input of the next, plain text as the universal interface.

## The five-layer context hierarchy

Every file an agent reads belongs to one layer, and the layer decides when it loads: a global identity/map file ("where am I?"), task routing ("where do I go?"), a per-stage contract of Inputs/Process/Outputs ("what do I do?"), stable reference material ("what rules apply?"), and per-run working artifacts ("what am I working with?"). Scoping context this way keeps each stage in the few-thousand-token range where models perform best, instead of a 30k–50k-token monolithic prompt of mostly-irrelevant context.

## Glass-box by construction

ICM's most cited theoretical move: interpretability becomes a property of the architecture, not a feature bolted on. Every intermediate artifact is a plain-text file a human can open, read, and edit between stages — so the artifacts *are* the log, and `ls` is the dashboard. The pattern fits sequential, reviewable, repeatable workflows; the paper is explicit that real-time multi-agent collaboration, high-concurrency systems, and complex automated branching are better served by conventional frameworks.

## Relationship to the standards

ICM is the workspace-level cousin of the [agent-readable web standards](/wiki/agent-readable-web-standards.md): both are *file-native context delivery*. [AGENTS.md](/docs/agents-md.md) asks how a repository tells an agent what to do; ICM asks how a workspace does. It has a working MIT-licensed reference implementation and named academic adopters including the University of Edinburgh's Neuropolitics Lab, ICR Research, and the Academy of International Affairs in Bonn.

---

## Backlinks

Pages that link here:

- [AGENTS.md](/wiki/agents-md.md) — An open Markdown format at a repository's root carrying the build, test, style, and security context a coding agent needs.
- [Agent-Readable Web Standards](/wiki/agent-readable-web-standards.md) — A family of plain-text conventions at well-known paths that let machine readers understand how a site or repository wants to be read — the protocol surface of the Independent Internet.
- [Compounding Knowledge vs. Retrieval](/wiki/compounding-knowledge-vs-rag.md) — Why an LLM-maintained wiki accumulates understanding where retrieval-augmented generation re-derives it on every query.
- [David McDermott](/wiki/david-mcdermott.md) — Co-author of the 2026 paper introducing Interpretable Context Methodology.
- [Glossary](/wiki/glossary.md) — Formal definitions of the key terms behind the Independent Internet and Web 4.0, as used across this wiki, with links to source material.
- [Interpretable Context Methodology: Folder Structure as Agent Architecture](/wiki/icm-paper.md) — Van Clief and McDermott's paper arguing that a well-organized folder hierarchy can replace a multi-agent framework: the folder structure is the orchestration.
- [Jake Van Clief](/wiki/jake-van-clief.md) — AI-systems engineer, content architect, and Marine Corps veteran; lead author of the ICM paper and originator of Interpretable Context Methodology.
- [LLM-Maintained Wiki](/wiki/llm-maintained-wiki.md) — A persistent, interlinked knowledge base that a language model builds and maintains — integrating each source once, rather than re-deriving answers from raw documents on every query.
- [LLM Wiki: A Pattern for Building Personal Knowledge Bases Using LLMs](/wiki/llm-wiki.md) — Karpathy's pattern for an LLM-maintained wiki: instead of re-deriving answers from raw documents on every query (RAG), the model integrates each source into a persistent, compounding knowledge base.
- [Independent Internet (Web 4.0)](/wiki/web4.md) — The web forgot who it was for. Web 4.0 — the Independent Internet — is a resilient, self-hosted internet people own and control. Explore the research.