ARES: Agentic Research & Evaluation Suite

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

Quick Start

Get ARES running in minutes - no API keys required!

Prerequisites

• Python 3.12 or higher
• Docker - For running code agents in containers
• 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.

Your First ARES "Agent"

No API keys needed!

Run the Hello World example to see the RL loop in action:

uv run -m examples.00_hello_world

This example uses:

• ✓ Local Docker containers (no cloud account needed)
• ✓ A mock LLM (no API keys needed)

You'll see how ARES treats code agent interactions as a reinforcement learning problem, with LLM requests as observations and LLM responses as actions.

Examples

ARES includes several examples that demonstrate different usage patterns:

1. Minimal Loop (Local Docker + Real LLM)

File: examples/01_minimal_loop.py What you'll need: Docker, Martian API key

# Set up your API key first (see Cloud Setup below)
uv run -m examples.01_minimal_loop

Shows the RL loop with a real LLM (via Martian API) and local Docker containers.

2. Local LLM (Fully Local)

File: examples/02_local_llm.py What you'll need: Docker

uv run -m examples.02_local_llm

Demonstrates running ARES completely locally using a local LLM (Qwen2.5-3B-Instruct). No cloud services required.

Cloud Setup (Optional)

For production use or larger-scale experiments, you can use cloud containers and API-based LLMs.

Option 1: Using Martian API for LLM Inference

1. Create an account at https://app.withmartian.com
2. Copy the example environment file: cp .env.example .env
3. Add your Martian API key: CHAT_COMPLETION_API_KEY=your_key_here

Option 2: Using Daytona for Cloud Containers

By default, ARES uses Daytona for container management. To set this up:

1. Create a Daytona account at https://www.daytona.io
2. Copy the example environment file: cp .env.example .env
3. Add your Daytona credentials:
   • DAYTONA_API_KEY=your_key_here
   • DAYTONA_API_URL=your_url_here

See .env.example for all available configuration options.

API Usage

ARES environments use an async version of the dm_env spec. Here's a complete example:

import asyncio

from ares.code_agents import mini_swe_agent
from ares.containers import docker  # Use local Docker, or import daytona for cloud
from ares.environments import swebench_env
from ares.llms import chat_completions_compatible


async def main():
    # Create an LLM client (requires CHAT_COMPLETION_API_KEY in .env)
    llm_client = chat_completions_compatible.ChatCompletionCompatibleLLMClient(
        model="openai/gpt-4o-mini"
    )

    # Load SWE-bench tasks
    all_tasks = swebench_env.swebench_verified_tasks()
    tasks = [all_tasks[0]]  # Run on only one task for now

    # Create environment with local Docker and MiniSWE agent
    async with swebench_env.SweBenchEnv(
        tasks=tasks,
        container_factory=docker.DockerContainer,  # Use local Docker
        code_agent_factory=mini_swe_agent.MiniSWECodeAgent,
    ) as env:
        # The RL loop
        ts = await env.reset()
        while not ts.last():
            # Environment sends observation (LLM request) to agent
            action = await llm_client(ts.observation)

            # Environment processes action (LLM response) and returns next state
            ts = await env.step(action)
            print(f"Step complete. Reward: {ts.reward}")


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

This example uses:

• Container backend: Local Docker (change to daytona.DaytonaContainer for cloud)
• LLM backend: Martian API (or any OpenAI-compatible API)
• Code agent: MiniSWE agent from the mini-swe-agent library