How to Choose Between a Deterministic Workflow, One Agent, and Multiple Agents

This article provides a practical framework for architects, engineers, and technical leaders to choose between deterministic workflows, single-agent systems, and multi-agent systems in enterprise AI. It argues that the default should be the lowest-complexity architecture, not an agentic one. Deterministic workflows are best for known paths, repeatability, low latency, and explainability. Single agents suit variable tasks within one domain and security boundary. Multiple agents are only justified when specialization, organizational boundaries, security isolation, parallel work, or context separation create measurable value. The article emphasizes hybrid designs where deterministic workflows handle state, policy, and audit, while agents handle model-driven judgment. It warns against over-engineering with multiple agents, which increases latency, failure modes, and operational complexity.