Jev has everyone talking about system-one models. In essence, system-one models are zero-shot classifiers. When it comes to classification and ranking tasks, Jev and its open-source alternatives are much faster and cheaper to run than exisiting frontier LLM models, making it a no brainer for incorporating them into my current workflows.

A survey of tools for tracing Claude Code - local log viewers, API interceptors, native OpenTelemetry, and eval platforms like Langfuse and Braintrust. The capture mechanism you pick decides what you can see, whether it scales to a team, and whether you can close an eval loop.

These are notes taken from a talk by Matt Pocock, where he shares his philosophy in designing agent skills. Matt is the author of the popular agent skill repository mattpocock/skills(154k stars).

Two months after Karpathy's LLM-wiki gist, there are 200+ implementations and a Google spec. Both vindicate the bet Wiki-Skills made — pure markdown, no database — and point to a handful of improvements now folded into the Wiki-Skills.

From WebFetch to Computer Use — the landscape of web tools available to CLI agents, the tradeoffs across speed, safety, and site coverage, and recommended combinations for common use cases.

What makes the Claude Code harness best-in-class? Let's dive into its leaked source code to learn about the rigorous engineering within. Turns out, its memory management is remarkably similar to the wiki system described in my previous post.