Introducing Prime Sandboxes: MicroVM sandboxes purpose-built for RL training. Model training requires running tens of thousands of concurrent sandboxes, leading to complex and costly configuration. We built Prime Sandboxes for our own team. Today we're releasing them publicly.

Prime Sandboxes are available both as standalone infrastructure through our CLI/SDK and as part of our RL suite. Users can enjoy the following features: 1. Full VM fidelity 2. Elastic capacity at scale 3. First-class RL support 4. Bring your own environment 5. Competitive tier-free pricing With an architecture built for agentic training and pricing designed for tens of thousands of concurrent instances, they are the most cost-effective sandboxes available today.

Here are some tasks you can run with Prime Sandboxes: 1. Perform a training run 2. Generate synthetic data 3. Run an eval 4. Run a persistent, remote agent All accounts begin with a limit of 1,024 concurrent sandboxes, and teams that need more can contact us directly.

Our pricing is built for scale, with no subscriptions or minimum spend. For our launch through December 22, we’re proud to offer the most competitive pricing on the market for sandboxes, at a third of the cost of other large sandbox providers.

Getting started is simple:

In the near future, we will expand Prime Sandboxes to offer GPU microVMs, state snapshotting, sandbox forking, and shared persistent workspaces. This foundation will enable autonomous research loops that can explore, recover, and compound progress over time. If you would like to be a part of our roadmap, join our Sandbox Platform team:

Read our full blog post here

Visit our sandbox page

our sandboxes are like oxygen to me most critical infrastructure by the goat @damian_b and @a_kirillo 🙌

absolute beasts @damian_b @a_kirillo

team cooked!! @damian_b @a_kirillo 🔥

so cool!!

nice

lfgggggg

the 1k concurrent limit is a massive flex for rl

@ad0rnai RL training is now cheaper and easier for everyone. The eng team really cooked on this one!

awesome guys, congrats !!!!! This is much needed

love the announcement!

awesome stuff!

great ship, congratulations @damian_b , @a_kirillo and rest of the team

🧑‍🍳

Cheap concurrent sandboxes move the RL bottleneck from infra to the grader. Thousands of rollouts learn exactly what the reward check rewards: check only that the code runs and you train a model that makes code run. A founder eyeing RL now spends the week on that check, not VMs.

cook! 👨🏻‍🍳🔥

@leonardofed LFG 🙌🏻⚡️ i was waiting for this!

finally, infra for rl

Could the environment, reward code, and a failure trace travel together so anyone can rerun an RL experiment?

@willcb collabed with modal?

For RL rollouts, reset semantics matter as much as start time. Pin each run to an environment image and snapshot, and exclude external side effects from the next sample. Otherwise identical tasks can get different rewards because sandbox state drifted.

If you’re serious about building agents, you need solid sandboxes. Not only for training agents, we also need them for inference to serve the request... Pretty good time to announce these ..

can it be self-hosted?

RL开始盖楼了

The useful boundary here is failure containment: VM fidelity is great, but the operator question is what survives a crash and how fast a bad run can be revoked. Per-run quotas and a kill switch matter as much as concurrency.

@willcb holy

the config problem is really a comparability problem: a rollout on a drifted image gives you a reward you can't compare to the rest, so the curve starts measuring drift instead of the policy. pin the env by hash and log it per rollout, so a batch is only ever charted against the same hash.

gpu support soon?

this is interesting

能不能通，并确认 harbor 的 codex agent 读哪些环境变量。

Prime Sandboxes purpose-built for RL training with tens of thousands concurrent MicroVMs solves a real infra pain

Seeing the maze of sandbox configs, I kept misplacing tweaks; I'm building to capture each change as a post and surface it where RL engineers chat.