Introducing Ciri: A "Fractal Swarm" Agent built from scratch with Google ADK

[valkryhx]

Hi everyone! 👋

I wanted to share Ciri, a project I've been "Vibe Coding" recently. It is a modern AI Agent system built entirely from scratch using the Google ADK (Agent Development Kit).

My goal was to demonstrate the infinite potential of Google ADK by building a system that feels like a next-generation AI OS—lightweight, scalable, and intelligent.

(Click the image above or links below to see the full demo)

🚀 What makes Ciri special?

1. Fractal Swarm Architecture ("Agent Smith" Style)
Unlike traditional hierarchical swarms, Ciri implements a Fractal Agent Architecture.

• Identical Nodes: Every node runs the exact same code. There are no hardcoded "Managers" or "Workers".
• Dynamic Roles: A node becomes a "Leader" simply because it received a task; it becomes a "Worker" when it accepts a sub-task.
• Service Discovery: Uses a lightweight local registry (swarm_registry.db) for nodes to find each other and self-organize.

2. True Dynamic Skills
Ciri starts with minimal meta-tools and loads capabilities Just-in-Time.

• Hot-Pluggable: Python toolkits can be mounted/unmounted at runtime without restarting the agent.
• Lazy Loading: We prioritize the get_tools pattern to ensure heavy resources (like browsers) are only initialized when actually needed.

3. Real-Time Steering & Control
Built on Google ADK's callback system, Ciri implements a robust Steering Mechanism.

• Intervention: You can interrupt a running agent in real-time.
• Fail Fast: If a worker in the swarm goes off-track, the leader (or user) can kill the task immediately to save tokens and time.

4. Infinite Context
Ciri uses an Auto-Compactor sub-agent that works silently in the background. When the context window fills up, it intelligently summarizes the history (lossy compression), allowing the agent to run indefinitely without losing the "gist" of the conversation.

🛠️ Tech Stack

• Framework: Google ADK (Python)
• Intelligence: Google Gemini / DeepSeek / Claude (via LiteLLM)
• Interface: Modern Web UI + Geeky TUI (Terminal UI) showing Interleaved Thinking.

📺 Video Demos (Agent Team Swarm)

Here is the complete demo of the Agent Team cluster in action:

• Part 1: Initialization & Dispatch
• Part 2: Batch Processing
• Part 3: Execution & Result

I hope this project inspires you to explore what's possible with Google ADK! Any feedback or contributions are welcome.

👉 Repo: https://github.com/valkryhx/google_adk_agent
📚 Deep Dive: Full Implementation Guide

Let's explore the future of Agentic AI together! 🚀