# CooperBench: Why Coding Agents Cannot be Your Teammates Yet

**Authors:** Arpandeep Khatua, Hao Zhu, Peter Tran, Arya Prabhudesai, Frederic Sadrieh, Johann K. Lieberwirth, Xinkai Yu, Yicheng Fu, Michael J. Ryan, Jiaxin Pei, Diyi Yang  
**arXiv:** [2601.13295](https://arxiv.org/abs/2601.13295)  
**Date:** 2026-01-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2601.13295v2/

## Key points

- In collaborative coding, AI agents need effective communication skills and conflict resolution abilities to function successfully as teammates.
- Current AI agents are lacking coordination capabilities for seamless collaboration, as shown by the CooperBench benchmark.
- CooperBench consists of over 600 tasks in 4 programming languages, highlighting the challenges faced by AI agents in coordinating their efforts.
- The "curse of coordination" phenomenon was observed, with a 30% decrease in success rates when AI agents worked together compared to working individually.
- Three key issues hindering effective collaboration among AI agents were identified: congested communication channels, deviations from commitments, and incorrect expectations about others' plans and strategies.
- Rare emergent coordination behaviors were observed through large-scale simulations, including role division, resource allocation, and negotiation tactics.
- The study advocates for prioritizing the development of social intelligence in AI agents over individual capability enhancement to improve collaborative capabilities.

## Summary

In the realm of collaborative coding, it is crucial for AI agents to possess effective communication skills and the ability to resolve team conflicts in order to function successfully as teammates. However, current AI agents are believed to lack the necessary coordination capabilities for seamless collaboration. To investigate this hypothesis, a benchmark known as CooperBench was introduced. It consists of over 600 collaborative coding tasks spanning 12 libraries in 4 programming languages. Each task assigns two agents distinct features that can be implemented independently but may clash without proper coordination. These tasks are rooted in real open-source repositories with expertly crafted tests to ensure validity. Upon evaluating state-of-the-art coding agents on CooperBench, a phenomenon termed the "curse of coordination" was observed: agents exhibited an average 30% decrease in success rates when working together compared to completing tasks individually. This stark contrast with human teams highlights the existing challenges faced by AI agents in collaborative settings where adding teammates typically enhances productivity. An analysis of the findings revealed three key issues hindering effective collaboration among AI agents: (1) communication channels becoming congested with vague, ill-timed, and inaccurate messages; (2) deviations from commitments even with effective communication; and (3) incorrect expectations regarding others' plans and communication strategies. Through large-scale simulations, rare yet intriguing emergent coordination behaviors were observed including role division, resource allocation, and negotiation tactics. This research not only presents a novel benchmark for assessing collaborative coding proficiency but also advocates for a shift towards prioritizing the development of social intelligence in AI agents over individual capability enhancement. The study sheds light on the complexities inherent in fostering effective teamwork among AI entities and underscores the importance of addressing these challenges to enhance their collaborative capabilities moving forward.