Putting Jev in front of Claude Code to decide context
Notes attributed to a ChatGPT co-inventor describe using Jev to decide within 100 milliseconds which messages, tool outputs and files stay in context, replacing Claude Code's compaction.
OKAY this is insane. The guy who co-invented ChatGPT just dropped his notes on building coding agents. I read it twice and honestly this might kill the way we're all using Claude Code right now. What he said. Claude Code carries your ENTIRE chat history into every single turn. That's why it gets slow and expensive after a few messages. His idea is to put Jev in front of Claude so it decides in 100 milliseconds what actually needs to stay in context. Every message, every tool output, every file. It basically kills compaction. And this is the part that broke half the Jev demos on YouTube, he says model routing can actually cost MORE because the big model still re-reads everything anyway. I've already started thinking about how to wire this into the agency's Claude Code setup. At this point I'm pretty sure six months from now the way we're using Claude Code today is going to feel completely ancient. Full notes below, definitely worth reading.
This post quotes another post: x.com/2867689615/status/2101894250401271876