Sorry for the probably dumb question, I’m no expert in infra stuff: With the change of

agent-server

to work like a long-running server, from the previous assumption that it’s ephemeral/max 15 days of life or so, is the Cloud intended to run long-running boxes, or is that even doable? Like, • is it supposed to become able to run with e.g. 15k events? • could it give you something similar to VMs?

is it supposed to become able to run with e.g. 15k events?

Vasco and I did some performance passes to get conversations scaling better with # of events, and to fix some potential memory leaks and that kind of stuff. Don't doubt we missed a few things and we might need to tweak based on use case but it's definitely a consideration.

Or 60k events 😅 What do you mean by based on use case? Use case could be e.g.

"/goal solve all the open issues in this repo, then run /security-check skill, open issues and fix them all, then /another-skill and do the same. For all fixes, open PRs and follow the PR process you have in AGENTS.md; ask for at least three AI reviewers from the list of

openhands -t /codereview-roasted

,

gemini

that you have in your memory."

That could be a long conversation, but it seems a fairly usual kind of use that you may like to set-and-forget for a while. Then the regular use cases would include interactivity; that is, I may still want to be able to restore the conversation a week later, and send another prompt or seventeen, right?

Oh, I think I just mean all our optimizations were speculative, and we'll figure out what the real pain points are when something breaks. The long-lived agent server implies different usage patterns than we've considered before.