This is a wonderful and eye-opening community and I am glad to be here. I am working through my first detailed ICM agents, and I want to drop a few highlights, perhaps to be helpful, perhaps to get feedback. First, I chose an example that I knew well, so that I could judge the output fully. In my case that was seeding data from production into a new sandbox - I know old school SAAS salesforce. Second, this is 100% logical, detailed and has a right answer. Here is what I found. Claude really wants to show off all of its capabilities, and this makes it difficult to corral its capabilities into a logical path. In other words, I want Claude to stay within the md files, and follow exact instructions. This became notoriously difficult, as it seems to be pulled to discussing future steps like load order, or accessing the mcp, when the first stage is only about getting a list of objects/tables from a reference file. Yes, I got snooty with the AI. Yes, it was fun. And thanks to everyone who talked about Claude's memory because I found it difficult to have Claude stop saving to memory. Since I am dealing with larger data pulls, it also wanted to push that data into its .claude project location - so I had to be more specific about saving the data into output files than I thought. In other words, 'save the data output into the ftg-data-load.md 'file did not result in a list of records and fields. I have to ask it to specifically add all field information so that the file can be used alone. I have loved the path forward the past week or so. If the pace holds true, the absolutely positively critical points above will seem like basic thoughts in about two days.