Built an AI Agent at Work? Ask Yourself What It Can Reach Before You Trust It

Many people who have built a small AI agent at work know the same sequence. First comes an hour of feeling like a genius. Then comes a look at what the agent can actually reach, and a quick decision to turn it off. A Reddit user who works in operations, not engineering, described that exact experience this week and asked other non-developers how their own automations have held up.

The poster separated two kinds of risk. A broken automation that touches their own files is recoverable. A broken one that touches company data is a different problem. An agent can work perfectly in a test run and then behave unexpectedly on an ordinary Tuesday, when the builder is in meetings and nobody is watching. In that situation, the person who set it up is the one who has to explain what happened. As the poster put it, telling the CEO that it worked when they tested it is not a strong career move.

The questions in the post were practical. What are non-developers really running at work, and what does it do? What happened once they stopped checking on it? Did it keep saving time, get quietly retired, or turn into a second job of maintaining it? The poster also wanted to hear from people who made the jump from "it works" to being able to take a holiday without worrying.

That last question is where the useful detail tends to be. Experienced builders usually point to a few habits: giving the agent read-only access wherever possible, limiting it to the specific data it needs, logging every action it takes, requiring a human sign-off before anything writes to a shared system, and keeping a clear way to switch it off. None of these are new ideas from software operations, but they are easy to skip when an agent takes an afternoon to build.

The broader lesson is that creating an agent has become easy, while running one responsibly remains hard. The gap between a convincing demo and a dependable process is where most of the risk sits, and that gap often falls on people who were never trained to manage it. Before asking what an agent can do, it is worth deciding what it is allowed to touch.