Stop Counting Tokens, Start Making Them Count

As AI coding agents move to usage-based billing, the question shifts from “does it work?” to “what did it cost to get there?” The common reaction — shorter prompts and token-counting — optimizes the wrong thing.

This session reframes the problem around agent quality as the cheapest optimization. It covers the mental model that makes cost legible: how a model differs from its harness, why a prompt is only the visible tip of the context window, and why output is the expensive part. It then walks through the practical levers — model choice, context engineering, Research → Plan → Implement, and deterministic guardrails — that reduce waste while preserving quality.