Transcription
Okay, so I bought a whiteboard because no one was understanding why agents are a waste of time. I don't mean using agents, I mean building them.
So in order to make AI not stupid, you need to route it to the right place to be able to give it the right instructions, give it the right tools, and give it the right data. Currently, people have built frameworks to do that. Lang chain, the Anthropic Agent SDK, Semantic Kernel, all of those are different ways to route these problems.
People spend this time building with Python or C, all these crazy agent frameworks only to get replaced by a model update from one of the big guys. This isn't because AI it's moving too fast. It's actually because of something else. People building agents are often operating at the wrong abstraction layer. Or more importantly, they're missing the most basic thing in computer science. File trees.
You can map every single AI agent out there to a simple file tree. Imagine you have a workflow and you have tasks inside that workflow. Also might have multiple workflows. These are just folders. Okay, pick inside of that workflow. Task one has instructions which are prompts, tools, and data. Same for task two. And you might have subtasks with its own instructions, tools, and data.
One single coding agent from any of the big guys would handle all of the instructions and processes in there and also have MCP servers which can get outside data, be able to create sub-agents which can handle other tasks and create new structures inside of your folder. All of that without you having to code in Python or C like Semantic Kernel requires.
Best part is this isn't going to get replaced by an update if they make a feature. One of the bigger people to do some of this, guess what? You just condense that into a tool. If they condense your workflow into a feature update, guess what? It's now a subtask or a tool. You are operating on the wrong abstraction layer. If you think AI is moving too fast.