The AI agent mistake everyone makes
The best AI agent I've built has a 14-word job description. The worst one I've seen comes with a four-page manual.
Every team is launching agents right now. Most will fail within a quarter — and the reason is almost never the model.
One job. That's it.
The agents that actually ship share a pattern: they do one thing, they refuse everything else, and they pull from one data source. Not because the builders lacked ambition — because they had the discipline to say no.
Here's what that looks like in practice:
A working agent's job description fits on a Post-it: "Rewrites product descriptions for new-arrival SKUs in brand voice, max 80 words." It doesn't touch subject lines. It doesn't run performance reports. It doesn't answer pricing questions — even when the team asks.
A broken agent gets a mandate like "help the merch team move faster" and inherits every unresolved disagreement in the company.
Refusals are a feature, not a bug
One brand's service agent resolved 60% of tickets cleanly — and quietly invented shipping dates on the other 40%. The first number showed up in weekly dashboards. The second showed up as chargebacks two weeks later.
An agent that says "I don't have that, routing you to a human" builds more trust in one exchange than an agent that confidently answers everything builds in a month.
Boring data decisions are where projects die
If your agent reconciles Shopify, your help desk, your email platform, and a spreadsheet, it will choose the wrong source at the worst moment. Language models don't flag uncertainty — they just write the most confident sentence the data allows.
Pick one canonical source before you build anything. If your stakeholders hesitate when you ask which system wins, you don't have an agent problem yet. You have a data governance problem.
The real differentiator
Model quality is plateauing. Tooling is commodifying. The teams actually shipping production agents aren't the ones with the fanciest stack — they're the ones who decided early what their agent wouldn't do, and held that line.
Constraint isn't thinking small. It's what gets you out of pilot.