109 days ago · 6 min read1219 words · Tech · hide · Fediverse post

Dating the layers Most of what we call an agentic harness is a fossil. It is the shape a workaround took when an older model could not do something. The workaround hardened into architecture, got a name, and kept being defended long after the model learned the trick on its own. You can date a harness the way you date rock layers. Show me a codebase with a vector store, a ReAct parser, a DAG runner, and a 4,000-line instruction file and I can tell you which quarter it was designed in. The uncomfortable part is that none of these were mistakes. Each one was the right engineering answer to a real weakness in the model. The mistake is keeping them after the weakness is gone. A workaround that no longer works around anything is just a wall between the model and the task. Small context: from RAG to agentic search When the window was 4k or 8k tokens, you physically could not put all your documents in the prompt. So we built retrieval: split the documents into chunks, embed the chunks, embed…

Commenting on Bubbles has been removed. Why? If you want to talk about this post, write to the author, use the comments on their blog if it has them, or reply with a "RE:" post on your own blog. Suggestions? feedback@bubbles.town