Current industry practice delegates agent coordination to LLM inference: the model picks which agent runs next, what data gets passed, and when to move between workflow stages. This paper argues that approach is unsound for production systems. I bring together evidence from three lines of research: (1) context degradation, where LLM performance measurably declines as input length increases, even well below nominal context window limits; (2) instruction-following failures, where current models satisfy fewer than 30% of instructions in agentic scenarios; and (3) the track record of deterministic compilation and DAG-based orchestration architectures that decouple planning from execution. I argue that reliable multi-agent workflows require treating orchestration as a runtime systems problem, governed by state machines, typed contracts, and deterministic transition logic, not as a natural language understanding problem. I propose six design principles for deterministic agent orchestration and identify open challenges.