A decision rule that's served us well is: start with the simplest orchestration that solves the problem, then earn your way to an agent.

Many workflows initially labeled "agentic" turn out to be deterministic pipelines with tool calling or structured outputs. Introducing an autonomous loop too early increases latency, cost, observability challenges, and testing complexity without adding much value.

The moment you genuinely need dynamic planning, retries based on intermediate results, or multi-step reasoning, that's where an agent framework starts paying for itself. Nice breakdown of where that boundary actually is.