Linear's telemetry shows coding agents multiplying output while planning and coordination do not shrink.
I have a habit of turning repeated arguments with coding agents into software. If I tell an agent to stay inside five files often enough, eventually I stop asking nicely and build something that blocks the sixth. If I keep asking for validation...
Google's proprietary family of multimodal foundation models, from the original Gemini releases through the agent-oriented 3.5 line and the efficiency-focused 3.6 Flash.
A technical lesson on using typed specifications as an agent control plane for preserving intent, governing delegated work, and verifying effects when generated code is cheap.