PrimeIntellect-aiOfficial·7 Agent Skills Included

research-environments

Benchmark environments for training and evaluating language models

Provides ready-to-run benchmark environments for training and evaluating language models across coding, math, science, terminal use, and tool calling. Eliminates the manual work of building datasets, sandboxes, and scoring rubrics from scratch. Runs smoke tests and full evaluations with a single command, saving every result to disk for analysis.

113|35

↗

npx skills add PrimeIntellect-ai/research-environments --all -g -y

Available:

Instructs your AI agent on how to write, run, and maintain benchmark environments in this repository, including coding standards, testing rules, and keeping skills updated after fixes.

All Skills in This Repository (7)

Pure Emerald Level Indicators

📦 In Repo

PrimeIntellect-ai

websearch

Execute concurrent Google web searches via the Serper API.

open_webpage

Fetch URLs and extract complete plain text from HTML and PDF documents.

synthesize-task

Generate tiered general-agent task families with empirical pass-rate gating.

cleanup-tasks

Diagnose and repair broken general-agent task definitions failing validation.

edit

Replace an exact string in a file when it occurs exactly once.

env-sync-push

Publish changed local environment packages to the Prime Intellect Environments Hub.

env-sync-check

Compare local environment versions against the Prime Intellect Environments Hub.

Official

Basic

Frequently Asked Questions

FAQPage Schema

How to install research-environments?▼

Run `npx skills add PrimeIntellect-ai/research-environments --all -g -y` in your terminal to install all skills in this suite globally.

What are Prime Intellect research environments?▼

They are ready-made benchmark tasksets for training and evaluating language models on coding, math, science, terminal, and tool-use tasks.

How do I run a benchmark evaluation?▼

Install an environment with `uv pip install -e path/to/env`, then run `uv run eval <taskset-id>` from the repository root to score a model on its tasks.

Can I build my own custom benchmark environment?▼

Yes. Each environment is a small Python package exporting a Taskset, and the included evaluation skill guides your agent through smoke tests and full runs.

Do these environments work with sandboxed runtimes?▼

Yes. Environments can run rollouts locally, in Docker containers, or on remote Prime and Modal sandboxes with network access controls.