Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Folder structure as agent architecture. ICM replaces framework-level orchestration with filesystem structure.
| Date | Stars |
|---|---|
| 2026-07-31 | 925 |
| 2026-08-05 | 955 |
| 2026-08-06 | 955 |
| 2026-08-13 | 1007 |
| 2026-08-18 | 1048 |
| 2026-08-19 | 1055 |
| 2026-08-20 | 1064 |
| 2026-08-21 | 1073 |
| 2026-08-22 | 1078 |
| 2026-08-23 | 1082 |
| 2026-08-24 | 1084 |
| 2026-08-26 | 1087 |
| 2026-08-27 | 1091 |
| 2026-08-28 | 1098 |
| 2026-08-29 | 1102 |
| 2026-08-30 | 1106 |
| 2026-08-31 | 1115 |
| 2026-09-01 | 1122 |
| 2026-09-02 | 1128 |
| 2026-09-03 | 1132 |
| 2026-09-04 | 1137 |
| 2026-09-05 | 1142 |
| 2026-09-06 | 1146 |
| 2026-09-07 | 1148 |
| 2026-09-08 | 1151 |
| 2026-09-09 | 1155 |
| 2026-09-10 | 1164 |
| 2026-09-11 | 1168 |
| 2026-09-12 | 1169 |
| 2026-09-13 | 1171 |
| 2026-09-14 | 1178 |
| 2026-09-15 | 1182 |
| 2026-09-16 | 1188 |
| 2026-09-17 | 1203 |
| 2026-09-18 | 1206 |
| 2026-09-19 | 1210 |
| 2026-09-20 | 1215 |
Today
+5 stars today
This week
+44 stars this week
This month
+142 stars this month
Momentum
49.0
growth rate 3.76%/day
# Interpretable Context Methodology (ICM) Folder structure as agent architecture. Full research paper here https://arxiv.org/abs/2603.16021 ICM replaces framework-level orchestration with filesystem structure. Numbered folders represent stages. Markdown files carry the prompts and context that tell a single AI agent what role to play at each step. The result is a system where one agent, reading the right files at the right moment, does the work that would otherwise require a multi-agent framework. **Created by Jake Van Clief** --- ## Why This Exists There are genuinely good agentic frameworks available today. CrewAI, LangChain, AutoGen, and others handle multi-step orchestration, memory management, tool use, and error recovery. They work. But they work within their own structures, and adjusting those structures requires development work. Changing the order of steps, swapping a prompt, adding or removing a stage: these actions typically mean editing code, understanding abstractions, and redeploying. For practitioners whose workflows are sequential and need human review at each step, the control surface can be much simpler. ICM is built on an observation that is almost too simple to write down: 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 tells it what to do at each step. This is going backward before going forward. The principles that made Unix pipelines effective in the 1970s -- programs that do one thing, output of one becomes input of another, plain text as universal interface -- apply directly to AI agent orchestration today. ## Design Principles Five ideas, each borrowed from established practice. **One stage, one job.** Each stage handles a single step. A stage that researches does not also write. A stage that writes does not also build. This follows the Unix principle and Parnas's information-hiding criterion. **Plain text as the interface.** Stages communicate through markdown files. No binary formats, no database connections, no proprietary serialization. Any tool that can read a text file can participate. Any human who can open a text editor can inspect or modify any artifact. **Layered context loading.** Agents load only the context they need for the current stage. Less irrelevant context means better model performance. This is prevention rather than compression. **Every output is an edit surface.** The intermediate output of each stage is a file a human can open, read, edit, and save before the next stage runs. The system picks up whatever the human left there. **Configure the factory, not the product.** A workspace is set up once with the user's preferences, brand, style, and structural decisions. After that, each run of the pipeline produces a new deliverable using the same configuration. ## How It Works Agents read down five layers and stop when they have what they need. ``` Layer 0: CLAUDE.md "Where am I?" Always loaded (~800 tokens) Layer 1: CONTEXT.md "Where do I go?" Read on entry (~300 tokens) Layer 2: Stage CONTEXT.md "What do I do?" Read per-task (~200-500 tokens) Layer 3: Reference material "What rules apply?" Loaded selectively (varies) Layer 4: Working artifacts "What am I working with?" Loaded selectively (varies) ``` Layers 3 and 4 are both content the agent loads while executing a stage, but they represent different kinds of context. Layer 3 is reference material -- design systems, voice rules, build conventions, domain knowledge. These files are configured once during setup and stay the same across every run. They are the factory. Layer 4 is working artifacts -- previous stage output, user-provided source material, anything specific to this run. These change every time. The distinction matters beca
Excerpt of 17,948 characters
Read on GitHub
20
1
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:56cf4577bcd115d5, llm:description: 'Folder structure as agent architecture. ICM replaces framework-level orchestration with filesystem structure.' Language: Python. No topics or README provided.
matched fp:56cf4577bcd115d5, llm:description: 'Folder structure as agent architecture. ICM replaces framework-level orchestration with filesystem structure.' Language: Python. No topics or README provided.