CooperBench

Can AI agents work together as teammates? CooperBench is the first benchmark designed to measure how well AI agents can cooperate when handling individual tasks with potential conflicts.

We find that coordinating agents perform much worse than a single agent given the same total workload. This coordination deficit presents a fundamental barrier to deploying AI systems that can work alongside humans or other agents.

Installation

pip install cooperbench

For development:

git clone https://github.com/cooperbench/CooperBench.git
cd CooperBench
pip install -e ".[dev]"

Requirements

• Python 3.12+
• Modal account (for sandbox execution)
• Redis (for inter-agent communication in coop mode)

Setup

1. Modal: Sign up at modal.com and run modal setup
2. Redis: Run locally with docker run -p 6379:6379 redis:7 or use a cloud provider
3. LLM API keys: Set in .env file:

ANTHROPIC_API_KEY=your_key
OPENAI_API_KEY=your_key
GEMINI_API_KEY=your_key

Dataset

Download the benchmark dataset:

git clone https://huggingface.co/datasets/cooperbench/cooperbench dataset/

Quick Start

CLI

Run agents on a task:

# Run cooperative agents (2 agents, shared communication)
cooperbench run -n my-experiment -r llama_index_task -m gpt-4o

# Run solo agent (1 agent handling both features)
cooperbench run -n my-experiment -r llama_index_task -m gpt-4o --setting solo

# Evaluate results
cooperbench eval -n my-experiment

Python API

from cooperbench import run, evaluate

# Run agents
run(
    run_name="my-experiment",
    repo="llama_index_task",
    model_name="gpt-4o",
    setting="coop",  # or "solo"
)

# Evaluate patches
evaluate(run_name="my-experiment")

CLI Reference

cooperbench run

Run agents on benchmark tasks.

cooperbench run -n NAME [OPTIONS]

Option	Description	Default
-n, --name	Experiment name (required)	-
-r, --repo	Filter by repository	all
-t, --task	Filter by task ID	all
-f, --features	Feature pair (e.g., 1,2)	all pairs
-m, --model	LLM model	gemini/gemini-3-flash-preview
-a, --agent	Agent framework	mini_swe_agent
-c, --concurrency	Parallel tasks	20
--setting	coop or solo	coop
--redis	Redis URL	redis://localhost:6379
--git	Enable git collaboration	disabled
--no-messaging	Disable agent messaging	enabled
--force	Rerun existing results	skip

cooperbench eval

Evaluate completed runs.

cooperbench eval -n NAME [OPTIONS]

Option	Description	Default
-n, --name	Experiment name (required)	-
-r, --repo	Filter by repository	all
-t, --task	Filter by task ID	all
-f, --features	Feature pair (e.g., 1,2)	all pairs
-c, --concurrency	Parallel evaluations	10
--backend	modal, docker, or gcp	modal
--force	Re-evaluate existing	skip

Experiment Settings

Setting	Agents	Description
coop	2	Two agents with Redis messaging, each handles one feature
solo	1	Single agent handles both features sequentially

Dataset Structure

dataset/
  <repo_name>/
    task<id>/
      setup.sh          # Repository setup script
      run_tests.sh      # Test runner script
      feature1/
        feature.md      # Feature description
        feature.patch   # Golden implementation
        tests.patch     # Test cases
      feature2/
        ...

Output Structure

Results are saved to logs/:

logs/<run_name>/<repo>/task<id>/features_<i>_<j>/
  agent1/
    trajectory.json     # Full agent trajectory
    patch.diff          # Generated patch
  agent2/
    ...
  eval.json             # Evaluation results

Benchmark Statistics

Metric	Value
Tasks	652
Repositories	12
Languages	Python, TypeScript, Go, Rust

Key Findings

1. Agents perform worse together than alone — GPT-5 and Claude Sonnet 4.5 achieve only 25% success with two-agent cooperation, roughly 50% lower than when a single agent handles both tasks.

2. Communication reduces conflicts but not failures — Agents spend up to 20% of their budget on communication, reducing merge conflicts but not improving overall success.

3. Three capability gaps underlie coordination failures:

   • Expectation failures (42%) — agents fail to integrate partner state information
   • Communication failures (26%) — questions go unanswered, breaking decision loops
   • Commitment failures (32%) — agents break promises or make unverifiable claims

Development

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Run integration tests (requires Modal)
pytest tests/ -v --run-modal

# Lint
ruff check src/
ruff format src/

# Type check
mypy src/cooperbench/

Citation

@article{cooperbench2026,
  title={CooperBench: Why Coding 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={2026},
  url={https://arxiv.org/abs/2601.13295},
  note={*Equal contribution (Stanford) · †Equal contribution (SAP Labs)}
}

License

MIT