Sailboxes are our cloud-environment specifically designed for long-horizon agents. We have been piloting Sailboxes in beta for the past few months, and today we’re excited to announce they are generally available!
Why use Sailboxes? #
We believe Sailboxes are an ideal fit for long-horizon agents for a number of reasons:
Cost efficiency. Sailboxes are priced >70% lower than other providers and charge only for actual CPU, memory, and disk consumption. Since multi-turn agents spend much of their time blocked on I/O, agents hosted in Sailboxes are significantly more cost effective. Sailboxes also require no reservations: agents can consume as much or as little compute as they need, and you only pay for what they use.Uncapped lifetimes. Background agents are most valuable when they are able to retain weeks of context at a time. Sailboxes support this by having no runtime limits.Full machines. Unlike other providers, Sailboxes give each agent full kernel-isolated Linux VMs with independent disks, persistent state, Docker support, and local NVMe. Your agent has root access and behaves the same way in the cloud as it does locally.
| Provider | vCPU | Memory | Max runtime | Mem snapshots | Sleeps on inference | Docker in Docker | Start/resume |
|---|---|---|---|---|---|---|---|
| Sailboxes | $0.015/ | $0.008/ | No fixed limit | Yes | Yes | Yes | <3s |
| Modal | $0.071/ | $0.024/ | 24 hours | Alpha | No | Alpha | <500ms |
| E2B | $0.0504/ | $0.0162/ | 24 hours | Yes | No | Yes | <1s |
| Vercel | $0.128/ | $0.0212/ | 5 hours | No | No | Yes | <1s |
| Daytona | $0.0504/ | $0.0162/ | No fixed limit | Experimental | No | Yes | <500ms |
| Cloudflare | $0.072/ | $0.009/ | No fixed limit | No | No | No | <2s |
How do we achieve this? #
Sailboxes achieve their efficiency through live migrations: we are able to achieve much higher utilization on our underlying hardware than other providers by migrating VMs based on their live resource usage. Enabling performant live migrations across heterogenous hardware is a complex problem that requires several bespoke pieces of infrastructure, including a custom network proxy and a peer-to-peer page faulting mechanism, but makes our fleet significantly more efficient. These migrations occur a few times a day, take several seconds, and happen completely without the knowledge of the agent.
While these migrations mean that the occasional command incurs a few seconds of additional latency, the tradeoff is massive gains in cost-efficiency. Unlike human-in-the-loop workflows like chatbots where patience is limited, long-horizon agents can tolerate the occasional latency blip without issue. Sailboxes directly trade off this p99 latency for better economics and we believe this makes them the ideal platform for any long-horizon agent.
Use Cases #
Hosting long-horizon agents
Sailboxes are currently used to power a number of long-horizon agents. Quadrillion Labs uses Sailboxes to host cloud-agents for their autoresearch platform, Qualia Cloud. When you ask a Qualia agent to test a hypothesis, it instantiates a fresh Python kernel within a Sailbox in seconds and runs multiple experiments in parallel. They are able to coordinate the results of their experiments between agents using our NFS-mounted Volumes product. Our partnership allows researchers on Qualia Cloud to massively accelerate their workflows with great economics.
Running RL Rollouts
Sailboxes have the ability to automatically sleep while an agent is waiting for the result of a Sail inference call. This feature makes them a great fit for performing RL rollouts. Rollouts are one of the primary use cases of asynchronous Sail inference: there is no human in the loop actively waiting for a result, and they benefit greatly from the massive scale that Sail inference provides. Combining Sail inference with Sailboxes allows you to run rollouts at maximal efficiency, giving you the ability to run more rollouts per dollar.
We have already run tens of thousands of Sailboxes for RL rollouts in conjunction with our Tinker-compatible API, and anticipate many more. We also have the ability to host any Prime Intellect environment.
Devboxes
One of the goals of Sailboxes is to give your agents a full computer in the cloud. The corollary of this is that they are also a convenient place to do development work. We already use Sailboxes within Sail to accelerate our development, both by hosting personal coding agents within them and also through asynchronous bug finding and automated triaging. You are able to stay connected to your Sailbox even as it migrates across hosts.
Getting started #
Sailboxes are available to all users today. To get started:
-
Install the Sail CLI:
curl -fsSL https://cli.sailresearch.com/install.sh | sh -
Enter our interactive shell and create your first Sailbox:
-
Visit our
[documentation](https://docs.sailresearch.com/sailboxes-quickstart)to learn more.
If you have any questions, you can find us at our [Slack community](https://join.slack.com/t/sailresearchcrew/shared_invite/zt-41pdcym9j-UU0Ey~A~r6n2H0DQVQsQHQ)!