Prime Intellect Launches Verifiers v1 for Agentic RL

Prime Intellect launched Verifiers v1 this week, introducing a significant rewrite of its framework for agentic reinforcement learning (RL) training and evaluations. The new version, available under the verifiers.v1 namespace, enhances the ability to run coding agents with integrated tools, compaction, and subagents. This architectural overhaul rebuilds environments to support agentic workloads at scale.

Previously, Prime Intellect's environment stack bundled data, agent logic, and infrastructure. Verifiers v1 decouples these components into three distinct, composable pieces: tasksets, harnesses, and runtimes. A taskset defines the specific work, including data, tools, and scoring criteria. A harness is responsible for solving the task and generating a rollout, which can be a ReAct loop, a command-line interface (CLI) agent, or a custom implementation. The rollout then executes within a runtime, which can be either local or sandboxed. This modular design allows any taskset to be run with any compatible harness.

The framework's architecture centers on a verifiers-managed interception server. This server acts as a proxy between the agent's runtime and the inference server, forwarding requests and responses while recording traces, setting sampling parameters, and rewriting tool responses to help mitigate reward hacking during training. For scalability, each server multiplexes a fixed number of rollouts, defaulting to 32, with an elastic pool that scales based on observed concurrency. A client relays these requests, with an EvalClient functioning as a blind HTTP proxy during evaluation and a TrainClient wrapping renderers for faithful token-in RL training.

To accommodate different agent communication protocols, Verifiers v1 supports three "dialects" as of its release: OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. A dialect adapter normalizes these various wire formats into a canonical vf.types format, ensuring that scoring logic remains independent of the specific agent being tested. This allows for flexible and standardized evaluation of diverse agentic systems.