Prime Intellect released Verifiers v1 on July 10, shipping a ground-up rewrite of its open-source environment stack for agentic reinforcement learning and evaluation. The release, announced on Prime Intellect’s blog and covered by MarkTechPost, introduces a three-part architecture that decouples what an agent learns from how it solves tasks and where execution happens.
Tasksets, Harnesses, Runtimes
The core design change: Verifiers v0 bundled data, agent logic, and infrastructure into a single monolithic environment. V1 splits that into three composable pieces.
A taskset defines the work: the data, tools, and scoring criteria. A harness is the program that actually solves the task and produces a rollout. That can be a ReAct loop, a CLI-based coding agent like Codex or Terminus 2, or a custom implementation. A runtime handles where execution happens, whether locally via subprocess, in Docker, or in remote sandboxes like Prime Sandboxes.
The decoupling means any taskset can run under any compatible harness. According to Prime Intellect, the team ported Terminal Bench 2 into v1 with minimal code and matched Harbor’s performance on the same tasks. Harbor is the first fully supported third-party format, with NeMo Gym and OpenEnv in alpha support.
Interception Architecture
The technical piece that makes swappable harnesses work is an interception server that sits between the agent’s runtime and the inference backend. According to Prime Intellect’s documentation, this server proxies all requests, records agent traces on the fly, sets sampling parameters for harnesses that don’t expose those settings, and can rewrite server-side tool responses to mitigate reward hacking during training.
Each server multiplexes 32 rollouts by default. A pool scales elastically with observed concurrency. Three API dialects are supported at launch: OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. A dialect adapter normalizes each wire format into canonical types, so scoring logic stays independent of the agent being tested.
Trace Efficiency
V1 replaces v0’s prompt-completion pair storage with a directed acyclic graph (DAG) of message nodes. In v0, each turn stored repeated prompt-completion pairs, producing quadratic growth in trace size as turns increased. V1 stores each message once and links it to its predecessor, keeping trace size linear in the number of turns.
For long-horizon agentic training, that difference is substantial. Branching is also native in v1, handling compaction and subagent rollouts without assuming linear traces.
Training Integration
The framework ships with full prime-rl training support from day one. In an internal ablation study described in the blog post, GLM-4.5-Air trained on ScaleSWE across six H200 nodes over two days and evaluated on SWE-Bench-Verified, demonstrating stable agentic training at scale.
The v0 code path is now frozen and will not receive active maintenance. Prime Intellect plans to fully deprecate it in a future release.
Where Agent Evaluation Infrastructure Is Heading
Most agent frameworks focus on deployment: running agents, connecting tools, managing conversations. Verifiers v1 targets a different layer entirely: standardized training and evaluation infrastructure. The release signals a growing recognition that building agents and measuring whether they actually work are two distinct engineering problems. As agentic RL moves from research papers to production training runs, the tooling gap between deployment frameworks and evaluation/training frameworks is becoming its own category.
The open-source repository carries 34 open issues and 229 pull requests as of publication, suggesting active community development around the new architecture.