What is 3and_agents?

3And (三和, Sanhe) is a Python framework that replaces the usual single-AI-answers-one-question pattern with a three-agent team that divides tasks automatically. You type one request, and three specialized agents handle it in parallel: a dispatcher that reads the intent and routes the work, a research agent that handles market analysis and information gathering, and an executor that writes code, builds web pages, or creates small games. When both working agents finish, the dispatcher reviews the results and attaches a confidence score and quality rating before delivering the output. The README is written primarily in Chinese and frames the project at business users: examples include generating a sales landing page, writing a market research report on Chinese AI companies, producing a data-cleaning script, and building a playable HTML5 game, all from a single sentence of input. The architecture uses Redis Streams as the message bus between agents. There are eight named message channels covering task dispatch, acknowledgment, progress updates, completion, failure, subtasks, interrupts, and heartbeats. Each working agent breaks its assigned task into a directed graph of subtasks (up to 10 nodes) and distributes them to a pool of workers that each call the DeepSeek language model API. Task state is stored in an SQLite database. Completed files are served through a file browser running on a separate port. The system runs in two modes. A lightweight mode requires only Python and a DeepSeek API key, with Redis disabled and a Gradio web interface on port 7860. A full Docker mode adds the Redis message bus for concurrent multi-agent processing. Adding a new task type involves three steps: registering a routing rule in the config file, adding a prompt template, and adding the name to the web interface. The project includes 123 automated tests that the README reports as all passing. It is MIT-licensed.