Recently a lot of excellent community discussion around how to ICM. Having experienced rapid growth in my past, one of the hardest things to do is keep language and definitions aligned. If I say the Agent did this, where does your brain go? Lines get crossed because you have so many levels of understanding and perspective. Some people are coming from non-technical backgrounds and have learned through corporate ai marketing. Others coming from corporate environments where language is defined for them by the management and designers building their infrastructure. Mini cultures with their own language. One of the brilliant aspects of ICM is it's commonality with almost anyone on a computer. It's just a folder, from that foundation we can all get on board. What's difficult is what happens next. I had this conversation with claude during a planning session. Claude asked me did I want agents or roles? I responded back by did he mean an agent like claude or codex, which is really an LLM, a purpose built agentic piece of software that the market tells me is an agent that does stuff? Or a folder acting as an agent with a specific role. I didn't know what the underlying meaning implied, how to answer? So I asked here was the response. 1. Engine — the LLM brand (Claude, Codex, Kimi, Minimax). Raw cognition, rented capacity, the electrons. Interchangeable. 2. Agent — the ICM folder. Persistent shaped context: identity + rules + reference + memory + boundaries. A standing operator that exists on disk and endures across sessions. This is your "folders hold attention" — an agent is where attention lives. 1. Role — a lens/hat: a function performed at one step of a process. It exists only in the doing. Where attention points right now.

Everyone building agents hits the same wall: the system only really exists inside a context window, and the moment that window rolls over you rebuild it from scratch in a fresh chat. So I moved the entire system state onto disk and built a loop that keeps it consistent. The maps layer is the part I would point at first. Five product maps plus a workflow map, every line anchored to a path-and-line citation. A job never mutates a map. It emits a structured flag, and a separate interactive pass re-verifies that flag against the live tree before it is promoted to a map line. Isolating that verification is what stops the maps drifting into the exact thing they exist to catch. The payoff is a queryable index over the whole system that I actually trust, and that trust is most of the sanity. The loop runs on two clocks. A deterministic kickoff fires once per task: registers the active worker, regenerates the tree, assembles its context. Zero model judgement, identical every run. Independently, a continuous harvest loop runs on a thirty-minute cycle, routing each job's record into the right surface, retiring finished workers, and flagging where a folder's own contract came up thin. Those surfaces are flat markdown, and that is the point. LIVE_STATE.md is the active set. ISSUES.md, PENDING_MAP_UPDATES.md and CONTEXT_FEEDBACK.md sit alongside it, each a single-writer file with one job, committed and fully diffable. Every executor reads the folder map before it touches anything, so it grounds its working set and writes to the right path instead of inferring one. It only works because the loop can read its own state back on every pass, and that state is the filesystem, not a window that forgets. Full video walkthrough below. How have you all layered on top of your ICM setups? Curious what you've added, what extra routers or surfaces you run, anything specific to your own structure you think is worth sharing.