Practical overview
When it fits
Use it when an agent evaluation must represent changing information, realistic task conditions and extended multi-step reasoning rather than a fixed prompt-and-answer test.
Who it is for
- AI agent researchers
- Teams benchmarking autonomous agents
- Developers studying multi-step reasoning and adaptation
- Researchers prototyping reinforcement-learning evaluation environments
Documented uses
- Run individual agent scenarios from the CLI
- Benchmark agents against Gaia2 dynamic scenarios
- Compare adaptation as new information becomes available
- Explore scenarios through an optional web GUI
- Evaluate agents beyond static benchmark prompts
Requirements & limits
The repository is public and MIT-licensed. The documented package supports Python 3.8 or newer, with uv/uvx or pip installation paths. A GUI installation extra is available for interactive exploration.
- The minimal installation is intended for CLI benchmarking and scenario execution; the GUI extra adds interactive exploration and monitoring.
- Gaia2 is documented as 800 scenarios across 10 universes, providing a broader evaluation suite than an individual scenario run.
- LiteLLM integration supports multiple model providers, but provider credentials and configuration can be required.
- The documented dynamic-environment approach is Verified for evaluating adaptation; adding a custom scenario is Experimental because no extension procedure is supplied.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run a basic ARE scenario with the mock provider
Check ARE command access and run the image-file scenario using the mock provider.
Before you start
- Python 3.12 or higher
- uvx, included with uv
- An API key is optional for basic testing
Steps
Check that ARE commands are available:
uvx --from meta-agents-research-environments are-run --help.Run the scenario:
uvx --from meta-agents-research-environments are-run -s scenario_find_image_file -a default --provider mock.Review the output for environment initialization, agent actions, environment updates, and the scenario result.
Expected outcome
The scenario runs with mock inference and displays execution details and a success or failure result.
Guide limitations
- The mock provider returns fake inference calls, so the scenario runs but the task will fail.
- Using a specific model instead requires configuring the environment and providing provider credentials.
Install ARE locally for scenario and agent development
Set up the ARE source code and command-line tools for development without building the UI.
Before you start
- Python 3.12
- Git
- uv
Steps
Clone the repository: the command below.
git clone https://github.com/facebookresearch/meta-agents-research-environments.gitChange into the repository directory:
cd meta-agents-research-environments.Create and activate a virtual environment: the command below followed by
source .venv/bin/activate.uv venv -p 3.12Install the project in editable mode: the command below.
uv pip install -e .Check command availability with
are-run --helporare-benchmark --help.
Expected outcome
ARE is installed locally in editable mode, and the documented command-line tools can be checked.
Guide limitations
- This setup is for scenario, agent, or core development and does not include the additional UI source-build steps.
- The activation command shown is for Linux/macOS; the cited installation instructions give a separate Windows setting for the UI-build case.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 4 October 2026
- Original website title
- facebookresearch/meta-agents-research-environments · Official homepage title · retained in its source language