
November 23–27, 2026
CTO
Genfoundry
AI agents turn seemingly simple user interactions into multi-step execution graphs involving model calls, retrieval, tool use, retries, handoffs and continuously growing context. Each individual step may appear reasonable, yet the overall workflow can quietly consume far more tokens—and cost—than engineers expect. This session examines token consumption as an execution architecture problem, not simply a model-pricing problem. We will look at common sources of avoidable consumption in agentic systems, including repeated context propagation, retry amplification, repair loops, retrieval overfetch, oversized tool payloads, redundant agent execution and inappropriate model selection. The session will also explore a harder question: how do we distinguish genuinely avoidable token consumption from necessary computation? Using execution-level examples, we will discuss the evidence needed to diagnose waste and why potential savings should only be quantified when the underlying execution data supports the calculation. Attendees will leave with a practical framework for moving beyond aggregate token dashboards toward deeper execution analysis: visibility → attribution → diagnosis → remediation → verified savings.
See every session across the AI Agents Conference November 2026 line-up.
View Agenda