Process
Tested: we have run it ourselves. Planned: part of the design, not yet run.
-
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.
-
Success
We agree what an agent should be able to do.
-
Tasks
PlannedTasks 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.
-
Soundness
TestedWe 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.
-
Baseline
PlannedWe measure an agent on the tasks with Prime Intellect’s Verifiers.
-
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.
- Prime Intellect’s Verifiers
- Runs the tasks and scores them. Results: React Best Practices Dependabot Gitmoji
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.