DevDuck agents

A multi-agent system for Go programming assistance built with Google Agent Development Kit (ADK). This project features a coordinating agent (DevDuck) that manages two specialized sub-agents (Bob and Cerebras) for different programming tasks.

Architecture

The system consists of three main agents orchestrated by Docker Compose, which plays a primordial role in launching and coordinating all agent services:

🐙 Docker Compose Orchestration

Central Role: Docker Compose serves as the foundation for the entire multi-agent system
Service Orchestration: Manages the lifecycle of all three agents (DevDuck, Bob, and Cerebras)
Configuration Management: Defines agent prompts, model configurations, and service dependencies directly in the compose file
Network Coordination: Establishes secure inter-agent communication channels
Environment Management: Handles API keys, model parameters, and runtime configurations

Agent Components

🦆 DevDuck (Main Agent)

Role: Main development assistant and project coordinator
Model: Qwen3 (unsloth/qwen3-gguf:4B-UD-Q4_K_XL)
Capabilities: Routes requests to appropriate sub-agents based on user needs

👨‍💻 Bob Agent

Role: General development tasks and project coordination
Model: Qwen2.5 (ai/qwen2.5:latest)
Specialization: Go programming expert for understanding code, explaining concepts, and generating code snippets

🧠 Cerebras Agent

Role: Advanced computational tasks and complex problem-solving
Model: Llama-4 Scout (llama-4-scout-17b-16e-instruct)
Provider: Cerebras API
Specialization: Go programming expert for complex problem-solving scenarios

Features

Multi-agent coordination: Routing between specialized agents
Go programming expertise: All agents specialize in Go/Golang development
FastAPI web interface: RESTful API with web interface support
Docker containerization: Easy deployment with Docker Compose
Flexible model configuration: Support for multiple LLM providers (local and cloud)

Quick Start

Prerequisites

Docker Desktop 4.43.0+ or Docker Engine installed.
A laptop or workstation with a GPU (e.g., a MacBook) for running open models locally. If you don't have a GPU, you can alternatively use Docker Offload.
If you're using Docker Engine on Linux or Docker Desktop on Windows, ensure that the Docker Model Runner requirements are met (specifically that GPU support is enabled) and the necessary drivers are installed.
If you're using Docker Engine on Linux, ensure you have Docker Compose 2.38.1 or later installed.

Configuration

You need a Cerebras API Key: https://cloud.cerebras.ai/
Create a .env file with the following content:

CEREBRAS_API_KEY=<your_cerebras_api_key>
CEREBRAS_BASE_URL=https://api.cerebras.ai/v1
CEREBRAS_CHAT_MODEL=llama-4-scout-17b-16e-instruct

look at the .env.sample file

✋ All the prompts are defined in the 🐙 compose file

Start the services

docker compose up
# if you updated the code, use --build

The application will be available at http://0.0.0.0:8000

Usage

The agents can be accessed through the web interface or API endpoints.

Activate Token Streaming

You can try this:

Hello I'm Phil

Bob generate a gloang hello world program

Add a human struct with a hello method

Cerebras can you analyse and comment this code

Can you generate the tests

✋ For a public demo, stay simple, the above examples are working.

🎥 How to use the demo: https://youtu.be/WYB31bzfXnM

Routing Requests

General requests: Handled by DevDuck, who routes to appropriate sub-agents
Specific agent requests
- "I want to speak with Bob" → Routes to Bob agent
- "I want to speak with Cerebras" → Routes to Cerebras agent

Tips

If for any reason, you cannot go back from the Cerebras agent to the Bob agent, try this:

go back to devduck