MoMoA: Mixture of Mixture of Agents

Coordinate independent LLM experts to solve complex, long-running engineering tasks that can exceed the capabilities of single-agent loops.

---

MoMoA breaks large projects into sub-tasks and assigns them to dynamic "Work Phase Rooms." Within each room, two specialized experts—like a Creative Developer and a Conservative Senior Engineer—are forced to debate, review, and validate each other's work before reporting back to an Orchestrator.

To start a session, point the CLI at your project directory and describe the goal:

python3 python_cli.py "Refactor the authentication logic to use JWT instead of sessions" \
  --directory ./my-web-app \
  --output ./updates

If the Orchestrator encounters an unresolvable ambiguity, it will pause and prompt you for a "Human-in-the-Loop" response directly in your terminal:

----------------------------Question from the agent:----------------------------
I found two different ways to implement the API endpoint. Should I prioritize 
execution speed or memory efficiency for this specific module?
---------------------------------------------------------------------------------

Your answer: Prioritize execution speed; memory is not a bottleneck here.

---

MoMoA is an experimental architecture tuned for the Software Development Life Cycle (SDLC). It prioritizes consistency over speed by requiring multiple rounds of internal peer review and validation before any file change is finalized.

CLI Reference

Argument	Description	Default
positional_prompt	The primary task description for the agent.	(Required)
-d, --directory	The local path the agent should read and modify.	None
-o, --output	Where to save worklogs and the final result diffs.	agent_output
-a, --assumptions	Path to a text file containing rules the agent must obey.	assumptions.txt
-s, --serverAddress	The address of the running MoMoA server.	localhost:3007
--no-save	Display diffs and results without writing all files to disk.	False

Setup & Configuration

1. Environment: Create a .env file in the server directory with your GEMINI_API_KEY
2. Launch Server:

npm install
npm run dev

3. Ignore Rules: Create an .agentignore file in the root of the project folder if you plan to have MoMoA run against an existing project. This follows standard .gitignore syntax to prevent the agent from reading heavy dependencies (like node_modules) or sensitive secrets.
4. Launch Client: Ensure you have websocket-client installed via pip:

pip install websocket-client
python3 python_cli.py "Your prompt here" -d ./your-project

Adding Support for Jules and Stitch Tools

To enable MoMoA to use the Stitch and Jules Tools you must sign-up for Jules and / or Stitch, obtain API keys, and update your .env file (or environment variables):

• JULES_API_KEY can be obtained from Jules following these instructions.
• STITCH_API_KEY can be obtained from the Stitch settings page.

The Jules API doesn't currently support repo-less tasks, so you must also provide a GitHub access token that has access to a GitHub repository that is connected to Jules, and in which we can create a temporary branch that the Jules Tool will use to provide access to Jules:

• GITHUB_TOKEN can be obtained from GitHub Developer Settings.
• GITHUB_SCRATCHPAD_REPO (Eg. myusername/my-private-jules-scratchpad-repo).

Key Architecture Components

• The Orchestrator: Breaks the prompt into sub-tasks and reviews work phase reports.
• Work Phase Rooms: Specialized environments (Engineering, Planning, Documentation) with domain-specific tools.
• Experts: Personas with conflicting prompts (e.g., "Skeptical" vs "Creative") designed to catch logical errors through dissent.
• The Overseer: A background process that triggers every 15 minutes to unstick the agent if it enters a circular logic loop.

About this Project

Project Home Page: https://labs.google/code/experiments/momoa

Code Home: https://github.com/retomeier/momoa

Maintained by: Reto Meier

License

This project is licensed under the Apache 2 License - see the license.md file for details.