abhishek

building sandbox with ebpf capabilities

[abhishek @abhi8twt]

https://t.co/Yr045sDiwJ

[abhishek @abhi8twt]

Companies providing sandbox out of docker can have tough times ahead, I feel most of the companies not covering use case need some moat for being in a profit for longer term or otherwise be like @modal covering from inference to sandbox which is really hard to replace by mere company who spins up some of sandboxes or be like @devindevelopers from outposts to cognition having their own use cases top of it with their models

[abhishek @abhi8twt]

on an average we are going under 50ms for boot time of our vm, thats one of the benchmark for us. boot time is good we already ran 100k sandboxes with peak boot time goes up to 100ms not more than that. focusing more how we can optimize our pauses more.
@themuthiah

[alien gaming @alien_gaming13]

my parents vibecoded me

[Muthiah- Building AI infra for modern AI Era @themuthiah]

hey, i'm Muthiah, founder of https://t.co/tBcyedAdBM

we're building secure, disposable computers for AI agents. every agent deserves a computer of its own.

Starting my build in public journey to share what we're building, the challenges we face, and everything i learn along the way.

looking forward to connecting with founders, engineers, and builders working on cool things.

let's build together

[abhishek @abhi8twt]

we are gonna give sandbox + integration to slack + linear + PR + Issue review in our dashboard, cli , sdk like the entire agent environment from end to end. vs code, browser and next month will have our own harness. end to end ownership and usecases are also coming stay tuned

[abhishek @abhi8twt]

how many lanes you guys using in your xfs question to all the sandbox companies

[abhishek @abhi8twt]

@reasonmachines where you focusing mate, review loop or meta loop. also wanna know about sandbox you are using, and the agents compared to devin

[abhishek @abhi8twt]

@modal has sandbox and have their own, and now multi-node training where you have the entire control of your infra because they have their own infra, this is what you should know and think about. you ain't gonna pick yourself up if you relying on other infra if your main moat is infra itself.

[abhishek @abhi8twt]

watching so many posts related to sandbox generic sandbox is not a goal of any customer, the one who have already positioned them are doing well. Wrap your one of the best use cases around it be it RL env, coding agent with higher terminal / harness benchmark. seems absurd but you can't see the sandbox if other people already grabbing the trust and others are open sourcing all of the stuff.

[abhishek @abhi8twt]

@striver_79 using namespace std;

[abhishek @abhi8twt]

@abshkbh @modal Use firecracker and forget, I personally feel its about security and host, you dont want that your server or host cpu getting affected

[abhishek @abhi8twt]

@aniketmaurya No doubt its overkill, if you can use a simple cmd to spin up the sandbox, working on mac sandboxes nowadays although so it wont align either

[abhishek @abhi8twt]

sandbox companies are eventually increasing, so does demand. but generic sandbox companies won't go anywhere soon increase the observability part, decrease boot time, use yaml configs for your industry to launch sandbox ax from google is showing lot of promises.

[abhishek @abhi8twt]

@hanghuang_ crazy guys, specially devops/deployment layers needs to get improved with agents, its tough but you got it.

[abhishek @abhi8twt]

many sandbox companies are in the market, but all have generic stuff, will be creating the best one, i got firecracker sandbox already working, now focusing audit trails using openshell + ebpf , let's see how it goes. trying to make someting that enterprise can use

[abhishek @abhi8twt]

People are using computesdk benchmarks way too serious just for marketing, no one's talking about how good the sandbox can be if you are providing observability patterns inside the kernel, you can make such a good combination by using two sides of the world which is wiz io and e2b as simple as that

[abhishek @abhi8twt]

gpu and cpu both will have crazy demand, cpu market has lot of demand but lots of sellers, for gpu this is not the case, even if the demand is lesser than cpu. that is the entire point. agents run on gpu but where they run is cpu, cognition solving both of the problems at same time for devin

[abhishek @abhi8twt]

xdp ebpf layer has lot of growth in future, your packet tells you a lot packet dissector still taking only what kernel needs, but other than that packet tells a lot of information. security observation part will be big leap for next decade, mark my words

[abhishek @abhi8twt]

startups will increase, so do customer distribution, gone are the time when one industry will be covering most of the customers, that was technology constraints. take an example of building sandbox industry lots of player lots of customer distribution, (not saying you shouldnt work on optimization)

[abhishek @abhi8twt]

Ebpf can come into picture here, non deterministic pattern of agents can easily caught with xdp by enforcing what packets needs to send, also lsm can be enforced so no multiple steps are been done, most important part is ebpf sensor have dedicated agent which learns how human and agent interacts instead, and learn the difference

[Nikita Bier @nikitabier]

Bot detection & human verification will be one of the most urgent demands for businesses over the coming years. Agent swarms will suffocate every website and form; small companies and government websites are most vulnerable.

There is a huge gap in the market for this right now. When we looked at what offerings were in the market to use at X, there was not a single company that brought together all the latest technologies so we had to do it all in-house.