Prime Intellect Releases Verifiers v1: Integrated Tasksets, Harnesses, and Workflows for Agentic RL Training and Testing

Introducing Prime Intellect confirmations 0.2.0. Previewing the rewritten theme, posted under new verifiers.v1 namespace. Modern testing now uses coding agents with tools, integrations, and subagents. Ideally, v1 is rebuilding environments to run these agent workloads at scale.

What are the credentials v1?

First, consider what reinforcements are: Prime Intellect’s natural stack for agent-reinforced learning and testing. In the past, the environment combined its data, agent intelligence, and infrastructure together. In contrast, v1 splits that bundle into three composable chunks.

A a set of work describes the work: data, tools, and scores. A harnesses resolves the task and generates the output. That harness can be a ReAct loop, a CLI agent, or your own. Discharge and internal a time to workeither local or sandboxed. Because the pieces are separate, any work set works under any corresponding harness.

How Architecture Works?

With those pieces defined, the next question is how they interact. The middle piece is managed authentication partition server. It resides between the agent runtime and the inference server. In particular, it includes requests for, and responses from, consideration. Meanwhile, it records the trace, sets the sample parameters, and can rewrite the instrument responses. That rewrite helps reduce reward hacks during training.

At scale, each server replicates a fixed number of outputs, by default 32. The pool then scales elastically in line with the marked currency. The server is also owned by the client forwarding those requests. During the test, i EvalClient it works as a blind proxy for HTTP. During training, a TrainClient to fold renderers for reliable RL training.

Because the harness speaks differently dialectsValidators support three as of now. This is the End of OpenAI Discussion, OpenAI Answers, and Anthropic Messages. A dialect adapter canonicalizes each phone format vf.types. Therefore, your understanding of the score remains independent of the tested agent.

v0 vs v1: A Quick Comparison

These variables separate v1 and v0.

Use Cases with examples

Now that the architecture is clear, think about how teams use it. For example, you can use Nemotron 3 Ultra in Terminal-Bench 2 under Codex.

Similarly, groups can reuse The port data sets without rewriting the reward logic. Prime Intellect has installed Terminal Bench 2 into v1 with only a small section. In its internal audit, the verifiers compared the Harbor’s performance in similar operations. The port is the first fully supported third-party format; NeMo Gym and OpenEnv have alpha support.

On the training side, the same areas connect directly to prime-rl. In removing the height penalty, GLM-4.5-Air is trained on ScaleSWE on six H200 nodes. That run lasted two days and was tested on SWE-Bench-Verified, which shows the agency’s stable training.

A Set of Small Tasks and Presentations

Each run starts with a work set that defines the data and points, without any harness:

import verifiers.v1 as vf

class AdditionData(vf.TaskData):
    answer: int

class AdditionTask(vf.Task[AdditionData]):
    @vf.reward
    async def exact_match(self, trace: vf.Trace) -> float:
        return float(trace.last_reply == str(self.data.answer))

class AdditionTaskset(vf.Taskset[AdditionTask, vf.TasksetConfig]):
    def load(self) -> list[AdditionTask]:
        return [
            AdditionTask(
                AdditionData(idx=i, prompt=f"What is {i} + {i}?", answer=2 * i),
                self.config.task,
            )
            for i in range(100)
        ]

__all__ = ["AdditionTaskset"]

Any task set then runs under the selected harness via TOML and CLI:

model = "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B"

[taskset]
id = "primeintellect/terminal-bench-2"

[harness]
id = "codex"
version = "0.116.0"

uv run eval @ path/to/config.toml

Key Takeaways

• verifiers v1 divide the area into a a set of work (what harnesses (how), and a time to work (there).
• Managed certifications partition server proxies-harness-inference requests and tracking records.
• The row message graph trace replaces v0’s quadratic prompt-completion pairs, allowing for long-horizon training.
• It is fully operational main-rl training support; the legacy code method is now fixed.
• The port datasets and harnesses alike The Codex again Term 2 work out of the box.

Check it out Technical details. Also, feel free to follow us Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.

Need to work with us on developing your GitHub Repo OR Hug Face Page OR Product Release OR Webinar etc.? contact us