Hey folks, I stumbled across this yesterday. It looks very interesting. After spending a bit of time reading the docs I still have a few questions. As far as I understand it, the Agent Server can be used as the “backend” process of the operation, if I don’t want the agents to run on my machine. After it is installed, the agents are configured via a client application (which can be the CLI, the Agent Canvas or a Custom Client). If I run the Agent Server in a container, will the agents inside that server run in containers of their own? The docs state that the enterprise version can run each conversation inside a separate sandbox, if wished. Is this an enterprise only Feature? How does this compare to the “normal” Agent Server? Are all agents just running inside the same sandbox in that case? If this has already been answered in the documentation, feel free to send me the links to the relevant sections with a smug comment. Thanks in advance

Hey, welcome around! Agent Server can run multiple conversations, including in the same time; agents will be in the same runtime environment but both they and/or the client application can separate their projects in different worktrees. That’s enough to not mix, in practice. Not in different docker sandboxes, typically. There is actually a legacy repo (openhands/sandbox-server) which was also creating containers for each conversation, which I think is what you might be interested in, but it was a bit heavy and fairly difficult to maintain, and pretty much ran its course - though you can try or point your agent at it. Agent Canvas is the recommended client, not the old CLI, or of course, all this is Open Source, so you can always make your own custom client app. As you see, technically container isolation is not exactly enterprise-only, but in practice it kinda is. One other option is to run the Agent Server on a remote VM. That separates them from your file system for real. Docker isolation is not perfect isolation after all, as you may have seen out there, long-running agents are so good these days that they will find funny ways to snap out of that “isolation”. 😅 I would suggest to try it. I think in practice, Agents running from Automations are separated from each other, since they have their own workspace with a copy of all they need; and Agents you prompt manually might not be many enough to step on each others’ toes, idk 😅