So agent harnesses like Claude and Antigravity recently introduced 1M-token context windows, which was great, at first. It turns out that a context window is re-sent on every turn, so a near-full window burns through your usage limit — FAST. Bummer.
Tagged “Agentic-Coding”
Intro -> Next Level
Nearly two years ago, early agentic coding let us build an on-device agent — in just a week — that would have taken months. I knew then that software was moving to the next level — today we call it agentic coding. Here's where I'll write about what that's actually like.