Library
Every issue
The weekly issue is free and complete, and it tells you what changed in AI research that week and what it means for the things you are building.
PPO agents trained on live opponents outplay frontier LLMs by 28 points when speed matters
If you build agents that have to decide under a deadline, you already know the gap between thinking fast and thinking well. This week one team closed it by building rather than prompting: a self-improving research agent …
[October 12–18, 2026]Agent skill systems are splitting between retrieval and self-evolution
If you run agents that fetch skills from a library, you are choosing between two architectures that used to look like one. This week the field drew a hard line: one camp treats skills as retrieved documents, the other as…
[September 28–October 4, 2026]Reward-free search and kernel-level kill switches redraw the agent stack
You ship agents that improve themselves or you ship agents that stay safe, and until this week those two pipelines barely spoke. Today they are the same conversation. Three labs showed that an agent can edit its own harn…
[September 21–27, 2026]Harness distillation without the harness at runtime
You have spent the week watching agents get faster by thinking less at test time. That trick has a ceiling: eventually the scaffolding that steers them becomes the thing that slows them down. This week the field went aft…