ARES

ARES (Agentic Research and Evaluation Suite) is an RL-first framework for training and evaluating agents.

Prerequisites

• Python 3.12 or higher
• uv - Fast Python package installer and resolver

To install uv, follow the instructions at https://docs.astral.sh/uv/getting-started/installation/

Installation

For now, we recommend running ARES locally from this directory:

uv sync --all-groups

and you're ready to get started.

Alternatively, include it as a dependency in your own project's pyproject.toml using a relative path. PyPI installation will be coming soon.

Configuration

ARES requires API keys for various services. To get started:

1. Copy the example environment file: cp .env.example .env
2. Edit .env and fill in your API keys (see .env.example for required and optional variables)

Getting Started

ARES environments use an async version of the dm_env spec. Below is an example snippet of what this might look like in your code.

By default, containers are run in Daytona, so you will need to:

1. Create a daytona account at https://www.daytona.io
2. Create a .env with DAYTONA_API_KEY=... and DAYTONA_API_URL=... set with an API key generated from your account.

This example also makes use of Martian for API inference. Similarly, you will need to

1. Create an account at https://app.withmartian.com
2. Add CHAT_COMPLETION_API_KEY=... to your .env with a Martian API key.

Then, you can run the following example:

import asyncio

from ares.code_agents import llms
from ares.environments import swebench_env


async def main():
    agent = llms.ChatCompletionCompatibleLLMClient(model="openai/gpt-4.1-mini")
    all_tasks = swebench_env.swebench_verified_tasks()
    tasks = [all_tasks[0]]  # Run on only one task for now.

    async with swebench_env.SweBenchEnv(tasks=tasks) as env:
        ts = await env.reset()
        while not ts.last():
            # The agent takes the observation (LLM Request)
            # and returns an action (LLM Response).
            print(f"Observation: {ts.observation}")
            action = await agent(ts.observation)

            # The environment takes the action (LLM Response)
            # and returns the next LLM request, reward, and discount.
            print(f"Action: {action}")
            ts = await env.step(action)


if __name__ == "__main__":
    asyncio.run(main())