AI training platform Prime Intellect has released verifiers 0.2.0, along with a preview of the next-generation Verifiers v1 architecture. Verifiers is an open-source framework that creates, runs, and scores tasks for AI agents — useful for both capability evaluation and reinforcement learning training.
Prime Intellect also open-sourced the model training framework prime-rl. In simple terms, Verifiers defines the tasks, tools, and scoring rules, while prime-rl trains models based on the task results. Developers can download and deploy both tools themselves.
The company also runs the Environments Hub and Lab. The former is a place to share and download ready-made training environments; the latter offers managed training services. Developers can either self-host the full stack or use Prime Intellect's environments and compute platform.
The old Verifiers tied tasks and agent execution together. v1 splits them into three components: Taskset defines what to do, which tools are available, and how to score; Harness determines how the agent completes the task; Runtime decides whether the task runs locally, in Docker, or in a remote sandbox.
This means the same task can be used with different agents — like Codex, Kimi Code, or Terminus 2 — and executed locally, in Docker, or in a remote sandbox. Developers no longer need to rewrite tasks and scoring rules every time they switch agents or execution environments.
v1 also records sub-agent calls, context compression, and other branching processes, saving training-required token IDs and log probabilities. The new version is better suited for long tasks spanning hundreds of steps and can directly feed agent trajectories into reinforcement learning. Future 1.0.0 plans to add multi-agent environments and improve support for environment frameworks like OpenEnv, NeMo Gym, and OpenReward.