Prime Intellect has released verifiers v1, a rewritten core for its reinforcement learning environment stack. The update shifts the platform from bundled monoliths to three distinct, composable modules: tasksets, harnesses, and runtimes.
The modular architecture
A taskset defines the work. It contains the data, tools, and scoring logic. A harness executes the task, producing a rollout. This component can be a standard ReAct loop, a command-line agent, or custom code. The rollout runs within a runtime, either locally or inside a sandbox.
Decoupling these elements allows any taskset to run under any compatible harness. The system now supports OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. A dialect adapter normalises each wire format into a canonical structure, keeping scoring logic independent of the specific agent being tested.
How the interception server functions
The central component is the verifiers-managed interception server. It sits between the agent’s runtime and the inference server, proxying requests and responses. The server records the trace, sets sampling parameters, and can rewrite tool responses to mitigate reward hacks during training.
For scale, each server multiplexes a constant number of rollouts, defaulting to 32. An elastic pool scales based on observed concurrency. During evaluation, an EvalClient acts as a blind HTTP proxy. During training, a TrainClient wraps renderers to ensure faithful token-in reinforcement learning.
What it means
Developers can now mix and match components rather than being locked into a specific environment bundle. The interception server provides a standardised way to observe agent behaviour and prevent training shortcuts without rewriting the agent logic itself.
