What nobody tells you about building your first AI agent

Most AI agent guides start with “pick a model.” That gets the whole thing backwards. The model is the last thing you should choose — after you've figured out the job, the judgment criteria, and the exact workflow the agent needs to own.

The agents that actually work in production didn't start with a model. They started with a person who had a problem worth solving.

Start with the work, not the tool

Before you touch any AI infrastructure, you need to answer three questions: What is this agent's job? What does success look like? And what does the human who owns this work need to trust about it?

If you can't answer those three questions clearly, you're not ready to build an agent. You're building a chatbot with extra steps.

The agents that fail — the ones that generate plausible-sounding nonsense, that require constant babysitting, that people eventually abandon — almost always fail at the same point: they were never clear about what job they were supposed to do.

The 72-hour test

Every agent we deploy goes through what we call the 72-hour test. We run it for three days without checking its output. If it produces work we'd throw away, it fails. If it produces work we'd actually use, it passes.

Most agents fail this test on the first try. That's fine. The test is what tells you what's broken — usually a missing piece of context, a boundary that wasn't defined, or a workflow that wasn't explicit enough.

The agents that pass are the ones worth keeping. Everything else is a prototype.