The most underrated skill in agentic AI: knowing when not to use an agent

The fastest agentic projects I've shipped are the ones where I took the agent out of half the steps.

There's a reflex right now to make everything an agent. A form gets submitted — send it to an agent. A file lands — wake an agent. It demos beautifully. Then it hits production and you're paying a language model to make a decision an if statement made correctly, deterministically, and for free.

An agent earns its place when the task genuinely needs judgment: ambiguous inputs, no fixed path, a decision that would otherwise sit in someone's inbox. For everything else — routing, validation, extraction against a known schema, anything with a correct answer — deterministic automation is faster, cheaper, and auditable. In a regulated environment, "auditable" isn't a nice-to-have.

The architecture I keep coming back to is boring on purpose: a thin deterministic backbone doing the predictable work, with agents placed only where reasoning is actually required. The agents do less, so they're easier to govern, cheaper to run, and far easier to explain to the person who has to sign off on the system.

Knowing when not to reach for an agent is a design skill, not a limitation. It's the same instinct as not putting a microcontroller where a switch will do.

Where are you seeing agents do work a plain rule should be doing?