Prime Intellect Releases Verifiers 0.2.0, Decouples Task Definition From Agent Execution

According to Beating, AI training platform Prime Intellect released Verifiers 0.2.0 and previewed its next-generation Verifiers v1 architecture. Verifiers is an open-source framework for testing and scoring AI agents, while the newly open-sourced prime-rl framework trains models based on task results. The v1 architecture decouples task definition, tool selection, and scoring rules (Taskset) from agent execution methods (Harness) and runtime environments (Runtime), allowing developers to reuse the same tasks across different agents like Codex, Kimi Code, and Terminus 2 without rewriting evaluation rules.

Developers can self-host both frameworks or use Prime Intellect's managed services via Environments Hub and Lab. The v1 design supports long-horizon tasks and directly captures agent execution traces for reinforcement learning training.