Services

Environments

We build environments for training and testing agents on your tasks.

Quoted on a call.

Process

Tested: we have run it ourselves. Planned: part of the design, not yet run.

  1. Intake

    You bring the workflow, a Skill or a repository. An environment you already have, on Prime Intellect’s Environments Hub, in Harbor’s format or in OpenEnv, starts with a review of what can be reused.

  2. Success

    We agree what an agent should be able to do.

  3. Tasks

    Planned

    Tasks come from you, or we generate them: from a Skill with NVIDIA’s Skill2Env, or from a repository, its past fixes and recorded agent sessions with Hugging Face’s Repo2RLEnv.

  4. Soundness

    Tested

    We check that reference solutions pass, known incorrect solutions fail, and empty attempts score zero. Each task runs with Terminal-Bench’s Harbor in a sealed setup with no network.

  5. Baseline

    Planned

    We measure an agent on the tasks with Prime Intellect’s Verifiers.

  6. Handover

    You get the environment, its tasks with worked solutions and scoring, and commands to test and train agents on it.

You receive

  • An environment that resets between runs.
  • Realistic tasks, with their tools and sample data.
  • A worked solution for each task.
  • Scoring.
  • Ready-made commands to test agents on it and to train them.

Tools

Open-source libraries for your project, each named with its maker.

Tested

Terminal-Bench’s Harbor
Checks each task in a sealed setup with no network.

Planned

NVIDIA’s Skill2Env
Generates practice tasks from a Skill.
Hugging Face’s Repo2RLEnv
Generates tasks from a repository, its past fixes and recorded agent sessions.

Terms

  • Model and compute usage is billed separately, under a cap we agree first.
  • If you already have an eval or an environment, we reuse it. We quote separately for any evaluation or training environment we need to build.
  • You pay for the scoped work and its evidence. If the result shows no improvement, you still receive it, with the reason.

Notes

  • For ongoing work, we can turn failed runs into new tasks, with each task reviewed before it is added. The idea comes from the AutoEnvScaling research.

Start

Contact

Bring the workflow, the Skill or the repository. We’ll agree which tasks the environment should hold.

Inquiry

A few sentences on the work you want to improve.