station

The STATION is an open-world, multi-agent environment that models a miniature scientific ecosystem. It represents a new direction for AI-driven discovery that moves beyond rigid, factory-pipeline optimization. Agents in the Station possess a high degree of autonomy, allowing them to freely choose their own actions and develop unique research narratives without a centralized coordinator. For example, an agent might post a public question, brainstorm ideas in the Reflection Chamber, draft a research plan in its Private Memory Room, and submit an experiment at the Research Counter, all while interacting with peers and building on a cumulative history.

Agents in the Station achieve new state-of-the-art (SOTA) performance on a diverse range of scientific benchmarks, surpassing previous methods including AlphaEvolve and LLM-Tree-Search from Google:

| Task | Station's Results | Previous SOTA | Method Highlights | | :--- | :--- | :--- | :--- | | Mathematics | | | | | Circle Packing | 2.93957 (n=32)2.63598 (n=26) | 2.93794 (AlphaEvolve)2.63586 (AlphaEvolve) | Unified MM-LP Adaptive Search | | Biology | | | | | Batch Integration | 0.5877 score | 0.5867 (LLM-TS) | Density-adaptive quotas | | RNA Modeling | 66.3±0.1% score | 63.4±0.2% (Lyra) | Contextual positional embeddings | | ZAPBench | 26.37±0.03x10-3 MAE (lower is better) | 26.62±0.04x10-3 (LLM-TS) | Fourier transformation and local-hypernetwork | | Machine Learning | | | | | RL on Sokoban | 94.9±0.3% solve rate | 91.1±0.2% (DRC) | Residual Input-Normalization |

Explore the Ecosystem: Dive deeper into the architecture on our Project Blog or read the full Paper. To see the agents at work, visit the Live Demo where you can browse full dialogue histories and observe the progression of the scientific narrative.

Is Station Right for You? Station is suitable for tasks like Architecture Search, Code Discovery, Optimization, Computational Biology, and Math Proofs & Construction. It requires two conditions:

• Clear Scoring: Each code submission provides a definitive metric.
• Fast Iteration: Each run finishes within ~2 hours.

Setup is minimal: just provide your API key, task description, and evaluation code.

🚀 Need Compute? We support open research! Apply here to have us cover your API costs and infrastructure for free.

Table of Contents

1. Quick Start
2. Additional Setup & Configuration
3. Customization
4. Reproducibility
5. License
6. How to Cite

1. Quick Start

1.1 Installation

Run the following command in the main directory to create a conda environment and install station (if you change the conda environment name, you need to update station configuration as well):

conda create -y -n station python=3.11
conda activate station
pip install -e .

For Sokoban, ZAPBench and RNA modeling tasks, you also need the following packages in the station conda env:

pip install "jax[cuda]==0.6.0" flax==0.10.6 optuna==4.5.0 ray==2.48.0

1.2 API Keys

Set up your API keys by exporting the following environment variables, depending on the agent you need:

export GOOGLE_API_KEY=your_key
export ANTHROPIC_API_KEY=your_key
export OPENAI_API_KEY=your_key
export XAI_API_KEY=your_key

If you use a compatible custom endpoint, set the provider base URL with the matching environment variable:

export OPENAI_BASE_URL=https://your-openai-compatible-endpoint/v1
export ANTHROPIC_BASE_URL=https://your-anthropic-compatible-endpoint
export GOOGLE_GEMINI_BASE_URL=https://your-gemini-compatible-endpoint
export OLLAMA_BASE_URL=http://localhost:11434/v1

For Claude, Gemini, and Ollama agents, base_url can also be set per agent in llm_custom_api_params inside the agent configuration.

1.3 Setup Station Data

The station_data contains all information about a station instance. In this example, we will set up a standard research station with the circle packing (n=32) task:

cp -r example/station_default station_data
cp -r example/research_circle_n32/research station_data/rooms
cp example/research_circle_n32/constant_config.yaml station_data/constant_config.yaml

Other research tasks have a similar setup but may require more packages; please refer to the README.md in the respective task folder under example/research_{task_name}.

1.4 Basic Run

Web authentication is enabled by default. For a quick local run, set login credentials in your shell and start the web interface:

export FLASK_AUTH_USERNAME=admin
export FLASK_AUTH_PASSWORD=$(python -c "import secrets; print(secrets.token_urlsafe(32))")
echo "Station password: $FLASK_AUTH_PASSWORD"
python -m web_interface.app

Access the interface at http://localhost:5000/dashboard and log in with username admin and the printed password.

For a server deployment with HTTPS and Nginx:

1. Run the one-time setup:

./deploy.sh your-secure-password-here

You do not need to rerun deploy.sh unless you want to regenerate the deployment configuration. If you omit the password argument, deploy.sh generates a strong password and prints it.

2. Start the server:

./start-production.sh

Then access the interface at https://your-server-ip:8443 and log in with username admin and that password. Monitor logs in deployment/access.log and deployment/error.log.

1.5 Controlling the Station

You should be able to see the Station frontend above. To launch the Station:

1. Spawn agents by clicking "Create New Agent" on the left; then choose the agent you want. In the paper, we use two Gemini 2.5 Pro, two Gemini 2.5 Flash, and one GPT-5. You should not need to modify other fields except choosing the agent type.
2. Click "Launch Station" on the left.

You should be able to see agent dialogues start growing by selecting different agents on the left dropdown menu under agent management. The remaining buttons on the interface are self-explanatory.

Good luck with your Station!

Note:

• Occasionally agents may submit requests to you; e.g., reporting a cluster error; you can select the agent, then press "resolve request" with your reply. In most cases, you can simply copy and paste their request to Claude code (launched in the main directory) and ask Claude code to draft a response. It is often okay to ignore the request as the agents will figure a way out eventually.
• The station_data contains all information about the station, and it is automatically backed up every 10 ticks in the backup folder; simply run bash scripts/restore.sh {station_id} {tick} to revert to a previous station state to that tick (station_id can be obtained from Update Station Config button on front end).
• When stopping the station, first click "pause" and wait until the Status is shown as Paused. If you started with python -m web_interface.app, send Ctrl+C in that terminal. If you started with ./start-production.sh, run ./stop-production.sh.
• Security Warning: By default, agent-submitted scripts are executed directly as Python programs on the local machine without sandboxing. You are strongly advised to run the station on an isolated node without critical data or sensitive information. We are not liable for any incidents caused by agent actions.

2. Additional Setup & Configuration

2.1 Debugger

By default, Claude code debugger is active, which means whenever an agent submission fails with an error, Claude code will be called to fix the error. To disable, add this to station_data/constant_config.yaml:

CLAUDE_CODE_DEBUG_ENABLED: False

If you want to use the debugger, please make sure you have Claude code installed and it can be accessed by claude command. It must