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 and David McDermott in their 2026 paper.
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: both are file-native context delivery. AGENTS.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.