Demos
Video

CoreWeave Sandboxes Demo

Play video

CoreWeave Sandboxes: isolated execution for AI at scale

Agentic AI doesn't just produce outputs—it executes code. CoreWeave Sandboxes contains that execution at scale, without adding a separate stack to your infrastructure. See how. 

1

00:00:05,833 --> 00:00:07,600

Hi my name is Deok

2

00:00:07,600 --> 00:00:08,533

I'm a product manager

3

00:00:08,533 --> 00:00:09,466

here at CoreWeave

4

00:00:09,900 --> 00:00:11,700

if you've been training agents

5

00:00:11,800 --> 00:00:15,433

running RL loops or model EVALs

6

00:00:15,433 --> 00:00:16,900

you've run into this challenge probably

7

00:00:16,900 --> 00:00:19,233

Code that came from the model

8

00:00:19,233 --> 00:00:22,066

has to run somewhere for you to verify it works well

9

00:00:22,566 --> 00:00:24,900

and you really don't want it touching your laptop

10

00:00:24,900 --> 00:00:27,800

or your training cluster that's what a sandbox is for

11

00:00:27,800 --> 00:00:30,533

an isolated environment spun up quickly

12

00:00:30,600 --> 00:00:34,166

that runs untrusted code human or model generated

13

00:00:34,566 --> 00:00:36,966

We're introducing CoreWeave Sandboxes

14

00:00:37,233 --> 00:00:39,366

there's basically two ways to use it

15

00:00:39,766 --> 00:00:43,700

You can use CoreWeave Managed Compute

16

00:00:43,700 --> 00:00:45,400

on a serverless option

17

00:00:45,466 --> 00:00:47,066

where you install the Weights and Biases

18

00:00:47,066 --> 00:00:50,100

SDK you log in and you can start running sandboxes

19

00:00:50,366 --> 00:00:52,866

there's no infra to set up and nothing to deploy

20

00:00:53,600 --> 00:00:56,400

There's also sandboxes for CKS and SUNK

21

00:00:56,866 --> 00:01:00,366

This infrastructure option is multi cluster by design

22

00:01:00,466 --> 00:01:03,300

it runs sandboxes on the clusters you already have

23

00:01:03,300 --> 00:01:05,266

on compute you're already paying for

24

00:01:05,566 --> 00:01:07,466

including idle CPU

25

00:01:07,466 --> 00:01:10,566

on the GPU nodes where you may be running SUNK

26

00:01:10,566 --> 00:01:13,333

which comes in very handy with capacity limitations

27

00:01:14,600 --> 00:01:18,266

Set up is easy for sandboxes through a simple CLI

28

00:01:18,300 --> 00:01:20,000

here you only need sandboxes

29

00:01:20,000 --> 00:01:21,700

admin IAM permissions

30

00:01:22,000 --> 00:01:24,700

and you create a profile that defines guardrails

31

00:01:24,700 --> 00:01:26,566

like networking policies for example

32

00:01:26,566 --> 00:01:29,900

to restrict ingress or internet egress

33

00:01:30,466 --> 00:01:32,833

you can also control namespace strategies

34

00:01:32,833 --> 00:01:34,166

and resource limits

35

00:01:34,866 --> 00:01:37,366

then you enable a runner on your cluster

36

00:01:37,566 --> 00:01:40,033

the runner is a CoreWeave managed workload

37

00:01:40,033 --> 00:01:43,266

which schedules each sandbox as a Kubernetes pod

38

00:01:43,366 --> 00:01:44,066

in your cluster

39

00:01:44,100 --> 00:01:45,700

Researchers in your org

40

00:01:45,766 --> 00:01:49,000

for example just do a pip install CW sandbox

41

00:01:49,033 --> 00:01:51,200

and they can immediately start running commands

42

00:01:51,200 --> 00:01:53,333

and code in sandboxes

43

00:01:53,366 --> 00:01:55,300

from their training scripts

44

00:01:55,300 --> 00:01:57,666

or whatever else they want to run it from

45

00:01:58,400 --> 00:01:59,466

from the SDK

46

00:01:59,500 --> 00:02:02,466

you can spin up thousands of sandboxes in parallel

47

00:02:02,466 --> 00:02:04,000

with a framework like veRL

48

00:02:04,000 --> 00:02:07,500

for example each rollout gets its own sandbox

49

00:02:07,500 --> 00:02:09,433

if you're running evals like SWE-bench

50

00:02:09,433 --> 00:02:11,566

which is a popular coding model benchmark

51

00:02:11,633 --> 00:02:14,966

you can spin up the hundreds of required containers

52

00:02:15,000 --> 00:02:18,100

and sandboxes to evaluate your model quickly

53

00:02:18,500 --> 00:02:20,233

So pick the path that

54

00:02:20,333 --> 00:02:21,866

best fits your needs

55

00:02:21,966 --> 00:02:23,233

and we can't wait to see

56

00:02:23,266 --> 00:02:24,533

what you build next