martian-ares 0.1.0

Agentic Research and Evaluation Suite

ARES: Agentic Research & Evaluation Suite

ARES is an RL-first framework for training and evaluating LLM agents, especially coding agents.

It is a modern gym: the environment layer powering RL research.

ARES treats LLMRequests as observations and LLMResponses as actions within the environment, so you can focus on training just the LLM - not the Code Agent surrounding it. The interface is entirely async, and supports scaling up to hundreds or thousands of parallel environments easily - check out example 3 to run this yourself.

Quick Start

Pre-requisites

• Python >= 3.12

Getting Started

Install with uv:

uv add martian-ares

ARES comes packaged with useful presets for different code agent & environment configurations. List them with:

uv run python -c "import ares; print(ares.list_presets())"

You can get started by using this minimal loop to run mini-swe-agent on SWE-bench Verified sequentially.

Note: to run this particular example you will need:

• Docker (with the daemon running)
• A Martian API key (see below)

import asyncio

import ares
from ares import llms

async def main():
    # This requires `CHAT_COMPLETION_API_KEY` to be set with a Martian API key--see below.
    agent = llms.ChatCompletionCompatibleLLMClient(model="openai/gpt-5-mini")

    async with ares.make("sbv-mswea") as env:
        ts = await env.reset()
        while not ts.last():
            action = await agent(ts.observation)   # observation = LLM request
            ts = await env.step(action)            # action = LLM response
            print(f"{action}\n{ts}")

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

To run the example above you'll need a Martian API key set in your .env file. To get a key:

1. Go to https://app.withmartian.com
2. on the Billing tab, add a payment method + top up some credits.
3. on the API Keys tab create an API key.
4. write CHAT_COMPLETION_API_KEY={your-key} in your .env

Alternatively, you can use another chat completions-compatible endpoint by setting both:

• CHAT_COMPLETION_API_BASE_URL
• CHAT_COMPLETION_API_KEY

Next Steps

1. Check out the examples
2. Read the docs to understand ARES and its key abstractions
3. Read our blog post about why ARES and what we hope to see