Agents that actually ship

Agents have a demo problem. They look great on a slide. They fall apart on a Tuesday.

Here's the lens I use when scoping an agent build with a client:

The 3-question test

1. What real task is this replacing? Not "intelligent assistant." A specific, named task someone does today.
2. What's the cost of being wrong once? This decides how much guardrail you need.
3. Who owns it on Monday? If there's no human on the hook, it will rot.

If we can't answer all three in plain English, the agent isn't ready to be built yet — it's still a wish.

Where I actually start

For most teams it's Zapier or Make for the orchestration layer, a single chat model in the middle, and tight scopes for v1 — one trigger, one outcome, no branching. Add complexity only after the boring version has been running for two weeks without anyone babysitting it.