When does an agentic design beat a deterministic pipeline?

Only when the path genuinely can’t be known in advance. If you already know that a task always needs, say, the logs and then the runbook, an agent “deciding” that dynamically at runtime is pure overhead — it adds latency, cost, and a new failure mode (a wrong decision) for zero benefit over just calling the two steps in order.

Agents earn their place on genuinely open-ended tasks, where what to do next depends on what the previous step actually found — the branching is real and data-dependent, not something you could have hardcoded.

My default architecture is a deterministic pipeline with model calls inserted at the specific stages that need language understanding or judgment, moving to a fully agentic loop only when I can point to real branching that can’t be enumerated ahead of time.