When an AI agent is the right tool — and when it isn't

The question we ask before building an agent isn't 'can an LLM do this' — it's 'does this task tolerate the failure modes of a probabilistic system'.

For workflows with clear success criteria and low tolerance for error, a deterministic pipeline with an LLM step is often the better architecture.

Agents earn their complexity when the task genuinely requires judgment across ambiguous, multi-step decisions.