jev was really exciting to us since it’s a reprisal of how we think a lot of systems with AI should be built!
think about it this way: when starting a project you don’t start with big behemoth k8s clusters (à la frontier models/ harnesses), you start with smaller units (jev) and iterate as needed
we learned A LOT when simple decisions from LLMs were the paradigm (which is what langgraph was built around), to see this type of model stack rank next to models of today (*on subset of tasks) makes us really excited for the future
Show more
Forking context is something you can do in a ton of different harnesses (claude, codex, now deepagents!) but it's something with very little guidance on when/how to use.
here's how we think about it!
managed deepagents entered into public beta last month - being 100p serious when I say it's been a big step change in how we operate internally
claims are a really cool exploration in what better memory for a coding agent could look like:
right now it looks like inline comments, ADR docs and skills, decisions scattered in loose markdown files...
But it's lossy and agents don't index it that well. Claims associate an agents decisions with the actual code it comes from, and openwiki can reconcile all its changes
It's also important that they're human readable which is what okf is for!
Show more
