cool-play-agent-teams 0.0.2

Role-driven multi-agent framework with strong typing and tool-based collaboration.

cool-play-agent-teams

Role-driven multi-agent orchestration framework built with strong typing and tool-only collaboration flow. Runtime model execution uses pydantic_ai with OpenAI-compatible endpoints.

Evaluation Snapshot

Recent SWE-bench snapshots are archived under docs/evaluations/swebench/. Current snapshots cover only the first 100 items from SWE-bench Verified, not the full benchmark.

Using glm-5,Temperature: 0.7,Top P:0.95.

Mode	Benchmark	Pass Rate	Passed	Failed	Mean Duration	Input Tokens	Cached Input	Output Tokens	Requests	Tool Calls	Report
Normal	SWE-bench Verified 100	72.0%	72	28	369.2s	60,265,198	58,214,976	451,537	2,432	2,484	HTML
Orchestration	SWE-bench Verified 100	73.0%	73	27	704.2s	103,016,077	95,659,776	1,886,195	6,026	7,171	HTML

Highlights:

• Orchestration currently reaches 73/100 on SWE-bench Verified 100, with 96 runs finishing in completed state and 4 ending in failed.
• Normal mode currently reaches 72/100 on SWE-bench Verified 100, with 97 runs finishing in completed state and 3 ending in failed.
• Token usage is reported directly in the table so model IO and tool activity can be compared without deriving cost assumptions.

Web Interface

Start the server with uv run agent-teams server start and open http://127.0.0.1:8000 in your browser. Use uv run agent-teams server restart to restart the managed server, and uv run agent-teams server stop --force to force stop it. The web UI now includes a language toggle beside the settings button so you can switch between English and Simplified Chinese in-page.

Frontend assets are now decoupled under frontend/dist and served by the backend.

Quick start

1) Install dependencies

Use the setup script for your platform, install from PyPI, or install directly with uv.

Windows:

.\setup.bat

Linux/macOS:

sh setup.sh

Install from PyPI:

pip install cool-play-agent-teams

Direct install:

uv sync --extra dev

2) help

agent-teams --help

# for evals
agent-teams-evals --help