Ask HN: How are you deciding when “agentic” AI is actually worth the complexity?

I’ve been watching more teams jump from simple LLM calls straight into “agents,” and I’m not convinced most of them need it.

The way I currently think about it is in three layers:

Single-shot generation – prompts, tools, outputs. Still underrated.

Structured workflows – fixed steps, guardrails, retries.

Agents – autonomy, memory, planning, and uncertainty.

In my experience, a lot of pain comes from skipping layer 2 and going straight to 3.

What’s helped me is asking:

Does this task actually benefit from autonomy?

Are failures cheap and reversible?

Would a dumb but explicit workflow solve 80% of it?

Curious how others decide:

What’s your “go/no-go” signal for agents?

Any heuristics you’ve learned the hard way?