The default way most teams build with LLMs makes the cost spiral inevitable. Raw context dumped into prompts. Whole HTML pages returned from tool calls. Cache-busting response shapes. Recursive agent loops that retry on cosmetic errors. The unit economics looks fine in the demo and catastrophic at scale.

This talk is a field report on the architectural decisions that determine whether your production AI agent costs five cents per query or one-tenth of a cent — an order of magnitude difference driven entirely by representation, shape, and cache strategy. I'll walk through the before/after on a real production MCP server where the architectural shift from raw HTML returns to structured representations cut response payloads by an order of magnitude — and what that meant for both latency and cost.

We'll cover: the representation choice that's hidden in every tool design and determines token efficiency more than prompt engineering ever will; why agent retry logic compounds cost geometrically and how to bound it; the cache strategy that works for agentic workflows (it's not the cache strategy that works for traditional web apps); and the audit pattern that lets you actually see where your tokens are going in a long-running agent run.

This isn't about prompt optimization or model choice. It's about the architectural primitives underneath both, which most teams aren't thinking about yet. If you're scaling a production AI product and your CFO is starting to ask uncomfortable questions about API spend, this talk is for you.