Æon Framework (Core)

English | Português

The Deterministic Runtime for Safety-Critical AI Agents with Autonomous Native Capabilities

🌟 Overview

Æon is a comprehensive, production-ready framework for building Neuro-Symbolic AI agents. Unlike stochastic-only systems, Æon combines the intuitive reasoning of LLMs (System 1) with the deterministic safety and control of code-level axioms (System 2).

It establishes a standard "Trust Stack" that enables agents to be Safety-Native, Protocol-First, and Extensible by Design. With deep integration of the Agent-to-Agent (A2A) and Model Context Protocol (MCP), Æon allows you to build interoperable agent ecosystems that can collaborate safely in high-stakes environments.

📋 What's New in v0.4.0 (ULTRA)

• 🔌 Autonomous Native Engine: Built-in support for browser automation (Playwright), persistent event-sourced memory (SQLite), and granular Trust Levels.
• 🏗️ Developer First CLI: Transform from scripts to projects with the new aeon command. Scaffold, run, and serve agents in seconds.
• 🚀 Declarative Runtime: Define agents via aeon.yaml and launch a full Gateway Server for production deployments.
• 🛡️ Enhanced Safety executive: Improved SIL-4 compliant axioms with TMR (Triple Modular Redundancy) reasoning for mission-critical reliability.
• 🔄 Deep Persistence: Event-sourced memory system that survives reboots and provides a complete audit trail of agent thoughts and actions.
• ⏰ Temporal Capabilities: Native scheduling for cron jobs and delayed tasks, enabling agents to act autonomously over time.

📋 What's New in v0.3.0 (ULTRA Phase)

• Routing Layer: Intelligent pattern-based message routing with 5 distinct strategies (Priority, Weighted, etc.).
• Gateway Layer: Centralized communication hub with session management and TTL support.
• Security Layer: Policy-based access control, AES encryption, and multi-provider authentication.
• Health Layer: Real-time system monitoring, metrics collection (Counter, Gauge, etc.), and diagnostics.

✨ Why Choose Æon?

• Deterministic Safety: Stop begging the model to be safe. Enforce safety at the runtime level with Axioms.
• Neuro-Symbolic Core: The perfect balance between LLM intuition and hard-coded rules.
• Protocol-First: Native support for A2A (Agent-to-Agent) and MCP (Model Context Protocol).
• Enterprise Ready: Built with observability, economics (cost tracking), and health monitoring from the ground up.
• Local-First & Private: Run entirely on your hardware with Ollama or connect to premium cloud providers.
• Stark visual Feedback: Terminal-native UI components for monitoring agent execution in real-time.

📦 Installation

Using UV (Recommended)

UV is the fastest way to manage Æon dependencies:

# Clone the repository
git clone https://github.com/richardsonlima/aeon-core.git
cd aeon-core

# Create environment and install
uv sync

Using pip

pip install aeon-core

🚀 Quick Start Examples

1. Developer Workflow (CLI)

From zero to agent in three commands:

# Initialize a new project
aeon init my-safety-agent

# Configure your model in aeon.yaml
# (Default: google/gemini-2.0-flash-001)

# Run a task interactively
aeon run "Check reactor thermal status"

# Start the production gateway
aeon serve --port 8000

2. Create a Safety-Native Agent (Code)

from aeon import Agent
from aeon.protocols import A2A, MCP

# Initialize the agent with the Trust Stack
agent = Agent(
    name="Sentinel",
    model="google/gemini-2.0-flash-001",
    protocols=[A2A(port=8000), MCP(servers=["industrial_tools.py"])]
)

# Define an Unbreakable Axiom (System 2)
@agent.axiom(on_violation="OVERRIDE")
def safety_limit(command: dict) -> bool | dict:
    """SAFETY RULE: Power output cannot exceed 100%."""
    if command.get("power", 0) > 100:
        return {"power": 100, "warning": "AXIOM_LIMIT_REACHED"}
    return True

if __name__ == "__main__":
    agent.start()

3. Autonomous Browser Workflow

from aeon import Agent
from aeon.core.config import TrustLevel

agent = Agent(name="Researcher", trust_level=TrustLevel.FULL)

async def main():
    # Agent can autonomously browse and remember
    response = await agent.run("Find the latest paper on SIL-4 safety and save the summary.")
    print(f"Agent Action: {response.last_thought}")

# Run via CLI: aeon run ...

🔌 Enhanced MCP (Model Context Protocol) v2.0

Æon now features a completely redesigned MCP implementation that provides robust, production-ready integration with external tools:

• Synapse Layer: Unified tool discovery and invocation.
• Standard Support: Full compliance with the latest MCP specification.
• Multi-Server: Connect to multiple MCP servers simultaneously (Stdio, SSE).
• Type Safety: Automatic parameter validation for tool calls.

📖 Architecture: The 16 Subsystems

Æon is organized into 4 distinct layers, each providing critical functionality for advanced agents:

1. CORE (System 1 & 2)

• Cortex: Neuro-reasoning via LLMs.
• Executive: Deterministic control via Axioms.
• Hive: Standardized communication (A2A).
• Synapse: Tool integration (MCP).

2. INTEGRATION

• Integrations: Multi-platform connectivity (Telegram, Discord, Slack).
• Extensions: Dynamic capability loading.
• Dialogue: Persistent, event-sourced conversation history.
• Dispatcher: Event-driven pub/sub architecture.
• Automation: Temporal task scheduling (Cron/Interval).

3. ADVANCED

• Observability: Life-cycle hooks and audit trails.
• Economics: Real-time token tracking and cost calculation.
• CLI: Premium developer interface.

4. ULTRA (Enterprise)

• Routing: High-performance message distribution.
• Gateway: Centralized session and transport management.
• Security: Authentication, authorization, and encryption.
• Health: System diagnostics and metrics.

🧪 Hello World: Industrial Overseer

from aeon import Agent
from aeon.protocols import A2A, MCP

controller = Agent(
    name="Reactor_Overseer_01",
    role="Industrial Automation Monitor",
    model="gemini-1.5-flash",
    protocols=[
        A2A(port=8000),
        MCP(servers=["mcp-server-industrial"])
    ]
)

@controller.axiom(on_violation="REJECT")
def enforce_safety(command: dict):
    # Any command attempting to disable cooling is rejected
    if command.get("action") == "DISABLE_COOLING":
        return False
    return True

if __name__ == "__main__":
    controller.start()

🖥 Terminal Output (Visual Feedback)

🚀 Æon Core v0.4.0-ULTRA initialized
├── 📡 A2A Server: Online at http://0.0.0.0:8000 (Unified Standard)
├── 🔌 MCP Client: Connected (4 tools loaded: read_sensor, adjust_valve...)
├── 🛡️ Axioms: 2 Active (enforce_safety, thermal_limit)
└── 🧠 Brain: Gemini-2.0-Flash (Ready)

🤝 Community & Support

• GitHub Issues: Report bugs or request features.
• Aeon Landing Page: Visit our landing page for deep dives.
• Contributing Guide: Learn how to join the mission.

📝 Citing this Project

If you use Æon in your research, please cite it as:

@software{richardsonlima-aeon-framework,
  author = {LIMA, Richardson Edson de},
  title = {Aeon Framework: The Neuro-Symbolic Runtime for Deterministic AI Agents},
  url = {https://github.com/richardsonlima/aeon-core},
  version = {0.4.0-ULTRA},
  year = {2026},
}

👨💻 Author

**Richardson Lima (Rick) **

• GitHub: richardsonlima
• LinkedIn: richardsonlima

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

---

Made with ❤️ for AI Safety by Richardson Lima.