Blog

Weekly roundups, field notes, and takes on agentic systems, evaluation, and applied AI. Curated from research and practice, written on Sundays.

RSS1 posts

LatestSep 6, 2026Roundup

Agents get measured, tokens get cheaper, output gets longer

A geometric law for how long an agent stays reliable, three Flash models in six weeks that cost less per token and more per task, and a cluster of papers that put the proof before the deployment.

8 min read43 sources2026-W36