Jake's ICM (March 2026) Google's OKF (June 2026) Now, Vercel's open-source Eve approach, just yesterday. There's an industry-wide convergence around "file-based" AI architectures happening right now. It's happening fast, maybe faster than I thought. Jake published ICM, a Unix-based methodology using folders and files, instead of code-heavy frameworks (AutoGen, LangChain, etc.), to create an architecture that runs sequential, HITL AI workflows. It's light, simple to build, simple to manage, and simple to share. It works well for most assisted work tasks and, while there's merit in other frameworks for handling more asynchronous, always-on, heavy automations, the industry seems to be waking up to the idea that for most jobs, a file-based system does the trick. I guess when Google published OKF, Vercel caught on to the wave and decided to publish Eve. In short, Vercel's Eve focuses on markdown files and TypeScript. An agent is a directory; you define the agent/directory.eve, add instruction.md files and TypeScript files for it to run. Eve compiles it into an app running on Vercel functions. Essentially, this is just Vercel introducing a form of ICM native to its platform. I read Jake's ICM paper (if you haven't, spend some time to do so), it's robust and plainly spells out what everyone else in the industry is starting to realize. For most workflows, simple is better. That isn't to say you can replace all your tools and processes with files. Even Jake himself states that for certain workflows there are better frameworks, but for most of our work now, and until AGI, this just works. What I take away from this: 1. Focus on the fundamentals and think in first principles as much as possible. Everyone is constantly putting out new tools, frameworks, etc. It's easy to become overwhelmed when you weigh everything the same and are constantly "rediscovering" things for the first time. When you understand the underlying principles, hundreds of "new" tools start to look like hundreds of variations of the same 5 things.