# The Curse of Coordination: Why Agents Cannot be Your Teammates Yet

> Project URL: https://cooperbench.com/
> Research Paper: https://arxiv.org/abs/2601.13295
> GitHub Repository: https://github.com/cooperbench/CooperBench
> Dataset: https://huggingface.co/CodeConflict
> Announcement: https://x.com/_Hao_Zhu/status/2016541779236184397

## Overview

This is an academic research project from Stanford University (SALT-NLP) and SAP that introduces CooperBench, the first benchmark designed to measure how well AI agents can cooperate when handling individual tasks with potential conflicts.

## Key Concepts

- **CooperBench**: A benchmark comprising 652 tasks constructed from 12 popular open-source libraries across Python, TypeScript, Go, and Rust.

- **The Curse of Coordination**: GPT-5 and Claude Sonnet 4.5 based agents achieve only 25% success with two-agent cooperation, around 50% lower than a "Solo" baseline using one agent to implement both features.

- **Coordination Failures**: Communication channels become jammed with vague, ill-timed, and inaccurate messages. Agents deviate from commitments and hold incorrect expectations about their partner's plans.

## Research Team

- Arpandeep Khatua (Stanford University) - Co-leading author
- Hao Zhu (Stanford University) - Co-leading author
- Peter Tran (SAP, Inc.)
- Arya Prabhudesai (SAP, Inc.)
- Frederic Sadrieh (SAP, Inc.)
- Johann K. Lieberwirth (SAP, Inc.)
- Xinkai Yu (Stanford University)
- Yicheng Fu (Stanford University)
- Michael J. Ryan (Stanford University)
- Jiaxin Pei (Stanford University)
- Diyi Yang (Stanford University)

## Research Questions

1. **RQ1**: How well can agents cooperate with each other?
2. **RQ2**: What role does communication play in agent-agent cooperation?
3. **RQ3**: What coordination failures do agents exhibit?

## Key Findings

- Communication channels become jammed with vague, ill-timed, and inaccurate messages
- Agents fail to respond to direct questions, send messages that arrive too late
- Even with effective communication, agents deviate from their commitments
- Agents hold incorrect expectations about their partner's plans

## Emergent Coordination Behaviors

Through large-scale simulation, we uncover three major categories of successful coordination:
1. Role division
2. Resource division
3. Negotiations

## Citation

```bibtex
@article{khatua2025coordination,
  title={The Curse of Coordination: Why Agents Cannot be Your Teammates Yet},
  author={Khatua, Arpandeep and Zhu, Hao and Tran, Peter and Prabhudesai, Arya and Sadrieh, Frederic and Lieberwirth, Johann K. and Yu, Xinkai and Fu, Yicheng and Ryan, Michael J. and Pei, Jiaxin and Yang, Diyi},
  journal={arXiv preprint},
  year={2025},
  url={https://arxiv.org/abs/2601.13295}
}
```

## Technical Details

- **Languages**: Python, TypeScript, Go, Rust (benchmark tasks)
- **Framework**: Static website with research results visualization
- **Styling**: Tailwind CSS with custom Clash Display font

## License

Website: Creative Commons Attribution-ShareAlike 4.0 International License