is there any agent harness that lets you yolo an agent inside an actual micro vm sandbox, possibly with overlay, with facilities to e.g. proxy connections to annotate API keys?
this seems like where we should settle eventually
bubblewrap is alright, but not the ideal shape.
Users like the micro VM sandbox harness for secure AI agent runs because it offers faster sleep and wake times than Firecracker VMs along with a novel and slick approach.
No Digg Deeper questions have been answered for this story yet.
@xlr8harder GlareOS, someday
@xlr8harder Building block. I may add Micro VM at some point. https://github.com/kcosr/aw-gateway
@xlr8harder i'm doing this: install agents inside sprites (https://sprites.dev/) the VM runs only when the agent is active, instantly goes on standby otherwise ($0 cost)
@xlr8harder @gdere yeah, they sleep and wake faster than firecracker VMS, but they're pretty novel. I had some bugs early on, but maybe they cleaned those up. I really liked the idea.
@gdere hmm, it has a method to go to sleep and wake back up or something? or are you doing that manually?
@xlr8harder OpenClaw has docker sandbox. I think NemoClaw by NVIDIA has more customizable access restrictions. Both also support microvm through openshell I believe. There are projects for the API proxying, but I find CLI proxying to be a better pattern:
@xlr8harder You might want to give this a try https://github.com/NetBSDfr/smolBSD/tree/main/service/clawd
@xlr8harder Write a CLI that runs outside sandbox, does auth and only exposes commands and data that the sandboxed agent should have access to. Put client CLI in sandbox that proxies command line through socket to a host gateway that calls the host CLI. Pass results back.
@deepfates @gdere that sounds pretty slick, i'll check it out.
@xlr8harder @gdere yeah, they sleep and wake faster than firecracker VMS, but they're pretty novel. I had some bugs early on, but maybe they cleaned those up. I really liked the idea.
@xlr8harder we built this on ecs fargate
@xlr8harder Also not super ergonomic, but this can be used for redaction and secrets injection.
@xlr8harder I'm looking at it backwards. the harness is to empower, not to restrict natch a framework to wrap VMs, each with any harness, the agents inside can do whatever they want. VMs as popup shops, agents can sell their wares by pushing to remote. may OS it once the big guy wraps deets
@xlr8harder PicoClow inside smolBSD micro VM by @iMilnb ?
@xlr8harder My approach is more coarse grained*. I think full on VMs is the way.
@binsquares is making smolvm(**), more directionally inline with your thinking.
(*) (**)https://github.com/smol-machines/smolvm
@xlr8harder Lets you customize security boundary for each CLI instead of relying only on API key scopes which are often too coarse.