This is actually a super interesting topic.
The core loop I’m seeing (at least as of literally just 09/22/2026) is a task execution framework that describes meta-level objects called Tasks (it can be injected by agents, users, pulled from Jira boards, etc), and then launches self-contained harnesses within VM’s to get the task done.
Of course, the granularity can shift; some people want VM’s for every tool call (which makes no sense to me, tool calls can be so arbitrary that you’re just going to be doing plumbing all the time it feels).
I’m aware of some companies trying to compete in the higher-level layer - really trying to make “task execution” formal (because you can compile anything to “task execution”), auditable, trying to make specs “translate down” in a sane way to code, etc.
I want to focus on the microVM aspect though. Not
gVisor or containers. Just VM’s. I’ve never actually
worked with something like Firecracker, but I would like to do
that. Just hack around and see what I can do. Certainly less
expensive than LLM inference. I’m not sure about the business use
case, but learning how people optimize VM’s for performance would
be a good challenge.