This session demonstrates developers and technocal leaders how to optimize their use of AI coding agents by prioritizing agent quality over raw token cost reduction. Covering LLM fundamentals, context window mechanics, the compounding error problem, and a practical toolkit of controls (model selection, prompt engineering, deterministic testing, agent configurations, and workflow phasing), the session equips attendees to get better results from fewer, smarter agent interactions.

This session explores how to maximize the return on investment of AI coding agents by shifting the focus from raw token cost reduction to agent quality improvement. It opens by framing the current state of agentic development as unsustainable "gambling" — sending under-specified agents repeatedly and hoping for good results — then introduces the compounding error problem to show why even small per-step accuracy gains dramatically improve end-to-end success rates. The session builds foundational knowledge of LLMs, context windows, and token mechanics (including phenomena like "lost in the middle" bias and context rot), before diving into a practical toolkit of quality and token controls: choosing the right model tier for each task, writing precise prompts with stop signals, dividing work into research-plan-implement phases, leveraging deterministic controls like tests and linters, and configuring agent instructions, custom agents, skills, MCP servers, and scoped prompt files. The core takeaway: instead of counting tokens, make every token count — because higher-quality agents naturally consume fewer tokens and deliver more value.