The VMs Powering Mobile Agents (Instinct, Claude Code)

Posted by RohanAdwankar 10 hours ago

Comments

Comment by alexellisuk 2 hours ago

I was hoping for a whole catalogue of providers - doesn't amp have their own offering here "Orbs"?

So, the interesting part for me was:

"Who runs the fleet Anthropic, its own E2B, rented, third party"

Like that's the only option, a bespoke platform or a rented SaaS for agent sandboxes.

Disclaimer: I am tooting my own horn here - we built SlicerVM.com (Firecracker + Apple Virtualization framework) so that people don't have to pick between those two options.

Especially when you have your own 1500-7000 USD MacBook right under your fingers.

For remote access, (and this happened by accident one night in January) we built superterm.dev - initially as a tmux manager, then as a way to run and monitor agents. Our team sits with 6-24 long running tasks with dictation available, great mobile experience, and can combine both for secure agents + high amounts of feedback.

If anyone's curious - Superterm has a CE edition - free for unlimited personal use. Slicer has a free trial and then a personal edition is available after that.

Comment by RohanAdwankar 1 minute ago

yes i havent tried out the rest of them, i learned about the space mainly just by poking around these 2 agents that I use. It seems very fun though since theres lot of optimization problems at play. Ill check out your one too would be cool to see what apples virtualization framework is like!

Comment by ushakov 31 minutes ago

Amp’s Orbs run on E2B.

E2B is fully open-source under Apache 2.0 and runs on GCP, AWS, Azure (preview) and locally on any Linux box that has /dev/kvm (Mac works via Colima, too).

source: https://github.com/e2b-dev/infra

Comment by 2078123478 56 minutes ago

[flagged]

Comment by hypendev 6 hours ago

It's always, firecracker all the way down!

Enjoyed the breakdown, honestly most of the things weren't really surprising.

What was surprising to me was that instinct was just "agent slapped on top of markdown files in git", especially after seeing people talk about it so much. I expected at least a database or a graph or something "smart", but alas, its just three markdowns in a trenchcoat.

Comment by RohanAdwankar 18 minutes ago

well to be fair this was just what i could reach inside the sandbox by grepping around if they are doing something else behind the api i wouldnt be able to see it

in guessing the most important thing to get right for their target usecase is the computer use and from the tools i saw for that the computers they provision were requestable by the subagents so probably something cool going on there

Comment by tensegrist 1 hour ago

this reminds me that i'm long overdue for a weekend spent studying firecracker, the userfaultfd trick [0], and whatever zeroboot does

[0]: https://codesandbox.io/blog/how-we-clone-a-running-vm-in-2-s...

Comment by RohanAdwankar 10 hours ago

I took a look inside the VMs that a couple of the agent platforms are using. TLDR is they use firecracker with memory in git. I'm a big fan of the memory in git part it would be cool to see how the git history looks after I use it for a longer period of time. This is my first blog, I'm trying to document things publically as I learn, next I'll look more closely at the providers they are using under the hood (eg. e2b) so feel free to give advice on what to look for!

Comment by htrp 34 minutes ago

How did you find their internal services ?

Comment by RohanAdwankar 25 minutes ago

accessed the vm through my terminal through a websocket: https://github.com/RohanAdwankar/ws-term

Comment by RohanAdwankar 22 minutes ago

so basically you convince the agent to run a script which allows you to get a “shell” and then from there i just grepped around and figured it out

Comment by thamizhan2611 7 hours ago

Do you have an invite to instinct? :)

Comment by RohanAdwankar 7 hours ago

yes i think everyone gets 10 did u want one?

Comment by user_7832 1 hour ago

Hi, if possible, could I get one too? You can contact me at esquire-zoom-jimmy@duck.com

Comment by thamizhan2611 6 hours ago

Yes, how can I reach out to you to get one :) thanks

Comment by RohanAdwankar 7 minutes ago

does dms via https://x.com/rohanadwankar work?

Comment by ushakov 4 hours ago

hey, i work at E2B, happy to answer any questions!

Comment by htrp 1 hour ago

super cool product, any thoughts on adding native llm support to your sandboxes ?

Comment by ushakov 24 minutes ago

my thoughts is that Firecracker might be not the best runtime for that, as it’s very minimal on devices it wants to support (famously no PCI out of the box)

Comment by RohanAdwankar 25 minutes ago

you guys are goated

Comment by FirstClassTree 4 hours ago

[dead]

Comment by quietraster 6 hours ago

great breakdown. did you notice meaningful differences in how the platforms handle filesystem persistence between sessions?

Comment by tensegrist 1 hour ago

    claude -p "${curl $url}\nreply only with a substantive comment in 'statement. question?' form" | dd conv=lcase

as a service

i wonder how hard it would be to write a decent classifier for individual comments like this (in practice you can just look at the account history which would make it an easy problem)