Essays and field notes on building agentic systems — the patterns we use ourselves, the mistakes we'd rather you avoid, and the parts of the craft that don't fit in a demo.
A short letter when we publish. No noise.
Most "digital employees" are chatbots with a headshot. Close the tab and they stop. The version worth building keeps the job open, acts overnight, and learns how you work across the tools you already live in.
Kimi K3 is the largest open-weights model anyone can download, and it runs on no machine you can buy. Deleting four-fifths of it made it fit. What that broke was not what anyone would predict.
Sonnet 5 is cheaper per token and costs ~2x more per task. The honest cost unit is now cost-per-job — and no leaderboard measures yours.
We'd optimised our agent harness across layers for months. Measuring 949 turns still surfaced ~90% per-turn overhead — a minor slip nothing was watching. The real lesson is observability, not the number.
From ChatGPT copy-paste to Cursor to Claude Code to planning to scheduled loops — a field map of how agentic work evolved, and why memory and iteration come before orchestration.
A May 2026 paper compiles whole agent workflows into small model weights — 87–98% of frontier quality, up to 462× cheaper per conversation. The mechanism, the catch, and why the build question just shifted.