I kept seeing versions of this question around the community: does ICM replace an automation platform like n8n, or does it sit on top of it?
That sent me down a rabbit hole. I built a fictional lead workflow, broke it in several different ways, and tested the handoff locally. What clicked for me was that we shouldn’t assign entire products to the 60/30/10 layers—we should classify the actual responsibilities.
The automation runtime can handle predictable events, API calls, retries, credentials, duplicate protection, and sending. The ICM workspace can make the current stage, relevant context, boundaries, working artifacts, and review path visible. Then the model handles the smaller piece that actually requires judgment.
The biggest lesson was that the handoff needs to work in both directions. “A lead arrived—go look” isn’t much of a contract. A proper task packet and a structured return status made missing information, ambiguity, retries, and human review much easier to route without guessing.
I put this video together partly to keep experimenting with Remotion, my paper puppet, and lip-syncing—but the community discussion gave me a genuinely useful system to test and explain. This isn’t meant as the final word—just my detailed walkthrough of what happened when I tried building the idea for real.
Here is the video link: