Are people actually running long-lived agents yet?or whether most people are still intentionally keeping agents short-lived because the runtime/reliability problems become too difficult.

Not copilots or request/response workflows but agents that:

survive restarts

continue tasks across sessions

maintain state over time

execute things reliably over hours/days

I’ve been thinking about this because it feels like once agents become long-running, the problem changes completely from prompting/model quality to runtime reliability.

For example:

after a crash/restart, what is the actual source of truth?

how do you know what already happened?

how do you avoid repeating side effects?

how much do you trust the agent’s own memory/reasoning after restart?

Most frameworks seem heavily focused on orchestration and tool use but I rarely see people talk about continuity, reconstructability or authoritative state over time.So whether people building serious agents are already hitting this problem like me and what architectures are actually holding up in practice.

[留言]

为什么值得关注

能改变理解方式，而不只是重复常识；符合当前抓取需求；它提供了新的理解或解释，而不只是表面观点

来源：reddit，领域：tech，保留分：0.67