🪁 Kite

Reliable AI Agents for Python

Fast • Safe • Simple • Powerful

---

Introduction

Kite is a lightweight Python framework for building production-grade AI agents.

Most agent frameworks rely heavily on prompt engineering for safety. Kite takes a different approach: it treats the LLM as an untrusted reasoning engine and enforces safety through code.

It provides a "kernel" that validates every action the agent proposes before it executes. This means you get strict control over permissions, blast radius, and failure handling.

Read the Architecture →

---

⚡ Quick Start

Installation

pip install kite-agent

Your First Agent

Here is a simple agent that proposes refunds but is supervised by the framework.

import asyncio
from kite import Kite

async def main():
    ai = Kite()

    # 1. Define Tools (The Execution Layer)
    # The only way the agent affects the world.
    def refund_order(order_id: str):
        print(f"💰 Refunding {order_id}...")
        return "Refunded"
        
    # Explicitly register the tool
    refund_tool = ai.create_tool(name="refund", func=refund_order)

    # 2. Define Policy (The Safety Layer)
    # Stop if it fails 3 times in a row.
    ai.circuit_breaker.config.failure_threshold = 3

    # 3. Create Agent (The Cognition Layer)
    agent = ai.create_agent(
        name="SupportBot",
        tools=[refund_tool],
        system_prompt="You are a support agent. Propose refunds if asked."
    )

    # 4. Run (The Kernel Supervision)
    # The framework validates the tool call before execution.
    result = await agent.run("Please refund order #999")
    print(result['response'])

if __name__ == "__main__":
    asyncio.run(main())

---

🎯 Core Philosophy

• Code over Prompts: Safety logic belongs in Python, not in English prompts.
• Whitelisting: Agents should only have access to tools you explicitly whitelist.
• Circuit Breakers: Prevent infinite loops and cascading API failures.
• Auditable: Every decision and action is traced.

---

⚡ The "Kernel" Pattern

In Kite, the agent doesn't execute code directly. It proposes actions to the Kernel.

# ❌ BAD: Logic in Prompt
# "Please check if the user is admin before deleting"
# Result: LLM might ignore this.

# ✅ GOOD: Logic in Code (Kite)
def delete_user(user_id: str):
    if not current_user.is_admin:
        raise SecurityError("Permission Denied")
    db.delete(user_id)

🚀 Key Features

• Multiple Reasoning Patterns: Support for ReAct, ReWOO, Tree-of-Thoughts.
• Production Safety: Built-in circuit breakers, idempotency keys, and shell access whitelists.
• Memory Systems: Vector RAG, Graph RAG, and persistent session memory.
• Multi-Provider: Switch between OpenAI, Anthropic, Groq, and Ollama purely via config.
• Human-in-the-loop: Pause execution for human approval on sensitive actions.

---

📦 Architecture Overview

Kite is modular. You can use the components independently.

kite/
├── agents/          # Reasoning engines (ReAct, Plan-Execute)
├── memory/          # Vector & Graph memory
├── safety/          # Circuit breakers & Guardrails
├── routing/         # Model routing (Fast vs Smart)
└── tools/           # Standard tool library

---

📈 Performance

• Startup: ~50ms (lazy loading)
• Memory: <100MB base footprint
• Throughput: Designed for high-concurrency async workloads.

---

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md.

---

📜 License

MIT License - see LICENSE for details.