September 2026 Summaries

Prime Intellect has launched Prime Sandboxes in general availability, offering Linux microVMs designed for large-scale agentic reinforcement learning workloads, with approximately 30 million sandboxes already created during early use. The service provides full VM fidelity, native Docker and Docker Compose support, hardware-level isolation, elastic capacity for tens of thousands of concurrent instances, and integration with Prime Intellect’s RL tools including verifiers, prime-rl, Hosted Training, and Prime Tunnels. The company argues that VMs avoid compatibility limitations of gVisor-based containers, allowing agents to train in environments closer to production systems while giving researchers greater control to prevent reward hacking. Users can deploy prebuilt or custom Docker-based environments through a familiar CLI and SDK workflow, drawing from a registry of more than 365,000 immutable environments for reproducible training and evaluation. Prime Sandboxes use usage-based, tier-free pricing and initially provide accounts with a 1,024-sandbox concurrency limit, while planned additions include GPU microVMs, state snapshots and forks, and persistent shared workspaces.