How do we approach agents that have to be deployed because there’s no way a user can run a filesystem on their computers, and we need to use tools like scripts? My guess is that I have to do 2 passes: 1. Pass 1: get the question and identify the “parameters” of the script 2. Use the response from the Llm to run a script, python or typescript function 3. Pass 2: use the output of the function to make it natural language format I would pass our identity + context + constraints in the Llm prompt to each of the passes. But doesn’t this looks very similar to complicated frameworks??

If you've been in this community for a minute, you've heard the term thrown around. ICM, Interpretable Context Methodology. Maybe you've seen it in Jake's lessons. Maybe you saw someone's folder structure screenshot and thought "that looks organized, but I have no idea what I'm looking at." 😅 I'm going to break it down the way I wish someone had broken it down for me. What ICM actually is (in plain English) 📂 ICM is a way of organizing your AI work, so the AI only sees what it needs to see, when it needs to see it. That's it. Instead of dumping everything into one massive prompt or letting a framework manage your context behind the scenes, you use your filesystem, folders, markdown files, plain text, as the architecture itself. Each folder is a stage. Each stage has one job. A CONTEXT.md file at the top tells the agent what this stage is, what inputs it expects, and what output it should produce. The agent walks into the room, reads the brief on the wall, does its job, and leaves. The next stage picks up the output. No frameworks. No LangChain. No AutoGen. Just folders and files. 🗂️ If you want to understand the philosophy behind why this works, where all of this leads, Jake lays the foundation here: 0.1: Where All Of This Leads - The Foundation Why it matters 🎯 Most people hitting a wall with AI aren't hitting a model limitation. They're hitting a context problem. The AI is trying to hold too much in its head at once. It forgets things. It contradicts itself. It hallucinates. It gets "lazy." 😴 That's not the AI being bad. That's you giving it a 47-page brief and asking it to stay sharp on page 43. ICM fixes that by isolating context. Each stage only loads what's relevant. The AI stays focused because you've structurally made it impossible for it to get distracted. The 60/30/10 rule (the lens behind it all) 🔍 This is the framework under the framework. Jake breaks every system into three layers: