Prime Intellect opens microVM sandboxes for agentic RL

Prime Intellect has opened general access to Prime Sandboxes, a managed fleet of Linux microVMs built for reinforcement learning workloads. Each sandbox has its own guest kernel, giving agents the system access needed to run Docker Compose, launch background services, modify file systems, and execute long-running tasks.

Agentic RL systems train and evaluate models through repeated task attempts called rollouts. Those rollouts often require thousands of isolated environments with different repositories, dependencies, and tools. Prime says its researchers and selected customers created about 30 million sandboxes during the product’s private rollout. General access is available through a CLI, an SDK, or Prime’s RL stack.

Guest kernels widen the sandbox

Prime Sandboxes do not use gVisor; instead, they use hardware-virtualized microVMs with their own guest kernels, in contrast to gVisor containers which use a user-space kernel. That model suits short-lived code execution, but agents can encounter unsupported system calls, restricted kernel features, and complications when running nested container tooling.

A microVM is a lightweight virtual machine optimized for rapid startup and high density. Prime places the isolation boundary at the hypervisor and gives each workload a guest kernel, which more closely matches an ordinary Linux host.

Prime cites a browser task in which file-system namespaces concealed a mock site’s backing data. The setup forced agents to complete the task through the browser instead of reading answer files directly. Guest-kernel control made those environment restrictions easier to configure and enforce.

Nested Docker workloads receive a similar benefit. Agents can use Docker Compose and related tooling inside the VM without relying on the partial kernel interface available in many container sandboxes.

A scheduler for mixed workloads

Common sandbox benchmarks measure how quickly a provider can launch many copies of one image. RL runs create a different scheduling problem: thousands of tasks may use separate repositories and environments, with several rollouts running for each task.

Prime’s scheduler caches images near available compute and favors hosts where the requested environment is already warm. The company advertises support for tens of thousands of concurrent VMs across the service. New accounts receive a limit of 1,024 concurrent sandboxes and can request a higher quota.

A container-shaped API

Developers submit Docker images through a familiar command-line interface, while Prime boots each image inside a microVM:

prime sandbox create python:3.11-slim
prime sandbox run <sandbox-id> -- python3 -c "print('hello')"
prime sandbox upload <sandbox-id> ./agent.py /root/agent.py

Public Docker Hub images work directly, while custom images can be pushed to Prime’s registry. Task images are stored immutably so previous runs can be reproduced against the same environment.

Prime’s RL integrations include its Verifiers framework and prime-rl. A companion service, Prime Tunnels, exposes a researcher’s local inference server through a public HTTPS endpoint, allowing parallel rollouts to call the server without custom network configuration.

Prime also offers a registry containing more than 365,000 prebuilt environments for open-source software engineering, terminal tasks, and agent evaluations. The company describes it as the largest environment catalog available from a sandbox provider.

The hourly bill

Prime says its launch rates are about one-third of those charged by other large sandbox providers. Billing is usage-based, with no subscription tiers or minimum commitments.

At those rates, a sandbox with 1 vCPU, 1 GiB of memory, and 32 GiB of disk costs $0.0389 per hour. Running 100 of that configuration concurrently costs $3.89 per hour. Prime says the promotional pricing runs through December 22; the launch announcement does not provide post-promotion rates.

Workloads that fit the model

Prime positions the service for workloads that combine Linux compatibility with high concurrency:

• RL training runs in which every rollout needs an isolated environment
• Synthetic data generation across many parallel workers
• Evaluations spanning different repositories, tools, and task images
• Persistent remote agents that maintain background processes

GPU and state controls come later

General access currently supports CPU workloads, while GPU microVMs, state snapshots, sandbox forking, and shared persistent workspaces remain on the roadmap.

Snapshots would preserve a VM’s state for later restoration. Forking would create multiple rollouts from the same saved state, allowing researchers to explore several actions from a promising point without repeating the preceding steps.

Production details remain unpublished

Prime’s launch material leaves several operational questions for teams evaluating the service:

• Available regions and data-residency options
• Startup latency under heterogeneous workloads
• Network egress pricing and bandwidth limits
• Maximum VM sizes and execution times
• Service-level commitments and support terms
• Hypervisor security controls and independent audits
• Standard pricing after the launch promotion

Teams currently assembling agent infrastructure from managed containers, hosted compute platforms, or custom Firecracker deployments gain a managed microVM option designed around RL scheduling. Its practical value will depend on measured startup performance, reliability, security controls, and pricing after the promotional period.