Coordinate a team of 30 AI coding agents in a single workspace
Running multiple AI coding tools simultaneously makes it difficult to track what each agent is doing and how far it has progressed. This tool creates a structure where agents exchange messages, record work status, and automatically report when they get stuck. Like a city mayor directing residents, a central coordinator manages the entire workflow.
Why agents do not forget
Existing AI agents often forget previous work content once a session ends. This system stores work state in Git-based hooks, allowing agents to maintain context even after a restart. It also tracks work items using structured data called 'biz', making it clear who is responsible for which task. The system is designed to reduce confusion and operate stably as the number of agents increases from 4-10 to 20-30.
Role division viewed through a city structure
The system calls the workspace a 'town', individual projects 'leagues', and the agents performing actual work 'polycats'. The 'mayor' acts as the main coordinator managing the overall workflow, while 'witness' and 'deacon' serve as monitors watching the health status of the agents. The 'refinery' is responsible for verifying and merging completed work, preventing agents from pushing directly to the main branch. This division of roles allows for systematic management of complex multi-agent workflows.
When stuck and when sessions break
If an agent gets stuck during work, an escalation feature automatically reports to a superior based on severity. The system also provides a 'seance' feature that finds records from previous sessions to restore context. You can ask and receive answers about decisions made in previous work, eliminating the need to re-read code from scratch in a new session. These features play a crucial role in minimizing context loss during long-term project execution.
Installation and getting started
On macOS, you can install gt, bd, and dolt together via Homebrew, while Linux and Windows require direct installation using Go. It is also possible to run in a container environment using Docker Compose. After installation, create the workspace with the gt install command and start necessary services with gt up.
Points to check before using this tool
This system supports various AI runtimes such as Claude Code, GitHub Copilot, and Codex, but requires configuration specific to each runtime. If you want a workflow using tmux, tmux 3.0 or higher must be installed. You can limit the number of concurrently running agents through scheduler settings to avoid API usage limits. Since understanding the complex structure may take time, it is recommended to start with small-scale projects to learn the concepts.
