Author outlines agentic workflow speed optimizations over 12 months

The author shares a mental model and practical lessons learned over twelve years of optimizing agentic workflows for speed. Key strategies include carefully analyzing inference speed components like TTFT and tok/s, accounting for hidden bottlenecks such as thinking budgets and inefficient tool calls, building specialized agent tooling, selecting appropriate hardware, and properly managing suspend and resume states in virtualized environments. Open article →