ceylonai-next 0.3.0

Ceylon: A Rust-based agent mesh framework for building local and distributed AI agent systems

Ceylon Python Bindings

Python bindings for Ceylon, a Rust-based agent mesh framework for building local and distributed AI agent systems.

Overview

Ceylon provides a unified API for creating agent-based systems that work seamlessly in both local (in-memory) and distributed (network-based) scenarios. The Python bindings allow you to build sophisticated agent systems using clean Python code while leveraging Rust's performance and safety.

Features

• 🤖 Custom Agents: Create agents with synchronous message handlers
• 🧠 LLM Integration: Built-in support for LLM agents (Ollama, OpenAI, etc.)
• ⚡ Async Support: Concurrent LLM operations with send_message_async()
• 🛠️ Actions/Tools: Define custom actions with automatic schema generation
• 🌐 Mesh Architecture: Local and distributed agent communication
• 📊 Metrics & Monitoring: Built-in metrics for performance, costs, and errors
• 🐍 Pythonic API: Fluent builder patterns and decorators

Installation

cd bindings/python
pip install -e .

Quick Start

Simple Agent

from ceylon import Agent, PyLocalMesh

class EchoAgent(Agent):
    def on_message(self, message, context=None):
        print(f"Received: {message}")
        return f"Echo: {message}"

# Create mesh and agent
mesh = PyLocalMesh("my_mesh")
agent = EchoAgent("echo")
mesh.add_agent(agent)

# Send message
mesh.send_to("echo", "Hello!")

LLM Agent (Synchronous)

from ceylon import LlmAgent

# Create and configure
agent = LlmAgent("assistant", "ollama::gemma3:latest")
agent.with_system_prompt("You are a helpful assistant.")
agent.with_temperature(0.7)
agent.with_max_tokens(100)
agent.build()

# Send message
response = agent.send_message("What is 2+2?")
print(response)

LLM Agent (Async)

import asyncio
from ceylon import LlmAgent

async def main():
    agent = LlmAgent("assistant", "ollama::gemma3:latest")
    agent.build()

    # Concurrent queries
    tasks = [
        agent.send_message_async("What is 2+2?"),
        agent.send_message_async("What is 3+3?"),
        agent.send_message_async("What is 5+5?"),
    ]

    responses = await asyncio.gather(*tasks)
    for response in responses:
        print(response)

asyncio.run(main())

Custom Actions

from ceylon import Agent

class CalculatorAgent(Agent):
    def __init__(self, name):
        super().__init__(name)

    @Agent.action(name="add")
    def add(self, a: int, b: int) -> int:
        """Add two numbers"""
        return a + b

    @Agent.action(name="multiply")
    def multiply(self, a: int, b: int) -> int:
        """Multiply two numbers"""
        return a * b

# Create agent
agent = CalculatorAgent("calc")

# Invoke actions
result = agent.tool_invoker.invoke("add", '{"a": 5, "b": 3}')
print(result)  # 8

Metrics and Monitoring

Ceylon includes built-in metrics collection for monitoring performance, costs, and errors:

import ceylonai_next as ceylon

# Run your agents...
# mesh.send_to("agent", "message")

# Get metrics snapshot
metrics = ceylon.get_metrics()

# Available metrics
print(f"Messages processed: {metrics['message_throughput']}")
print(f"Avg latency: {metrics['avg_message_latency_us']/1000:.2f} ms")
print(f"LLM tokens used: {metrics['total_llm_tokens']}")
print(f"LLM cost: ${metrics['total_llm_cost_us']/1_000_000:.4f}")
print(f"Memory hit rate: {metrics['memory_hits']/(metrics['memory_hits']+metrics['memory_misses'])*100:.1f}%")
print(f"Errors: {metrics['errors']}")

Key Metrics:

• message_throughput - Total messages processed
• avg_message_latency_us - Average message latency (microseconds)
• avg_agent_execution_time_us - Average agent execution time (microseconds)
• total_llm_tokens - Total LLM tokens consumed
• avg_llm_latency_us - Average LLM API latency (microseconds)
• total_llm_cost_us - Total LLM cost in micro-dollars ($1 = 1,000,000 μ$)
• memory_hits/memory_misses/memory_writes - Memory operation counts
• errors - Dictionary of error types and counts

See examples/README_METRICS.md for detailed examples.

Examples

Example scripts are located in the examples/ directory, and tests are in the tests/ directory.

Basic Examples

• examples/demo_simple_agent.py - Basic agent with synchronous message handling

  python examples/demo_simple_agent.py
  

• examples/demo_agent_mesh_local.py ⭐ NEW - Local mesh networking with multiple agents

  python examples/demo_agent_mesh_local.py
  

  Demonstrates:

  • Creating a local mesh network (PyLocalMesh)
  • Adding multiple custom agents to the mesh
  • Direct agent-to-agent messaging
  • Message routing patterns
  • Agent statistics tracking

• examples/demo_conversation.py - LLM agent conversation (synchronous)

  python examples/demo_conversation.py
  

• examples/demo_llm_mesh.py ⭐ NEW - LLM agents in mesh network

  python examples/demo_llm_mesh.py
  

  Demonstrates:

  • Multiple LlmAgents working together in PyLocalMesh
  • Specialized agents (coordinator, research, code assistant)
  • LlmMeshAgent wrapper pattern for mesh compatibility
  • Using Ollama Ministral-3:8b model
  • Agent-to-agent LLM communication

Async Examples

• examples/demo_async_llm.py ⭐ NEW - Concurrent LLM operations (recommended)

  python examples/demo_async_llm.py
  

  Demonstrates:

  • Concurrent queries with asyncio.gather()
  • Streaming responses with asyncio.as_completed()
  • Batch processing with concurrency control
  • Error handling in async contexts

• examples/demo_async_agent.py ✨ NEW - Async message handlers and actions

  python examples/demo_async_agent.py
  

  Demonstrates:

  • Async on_message() handlers
  • Async action execution
  • Thread-local event loop handling

Metrics Examples

• examples/metrics_quickstart.py ⚡ NEW - Quick start guide for metrics

  python examples/metrics_quickstart.py
  

  Demonstrates:

  • Basic metrics collection with get_metrics()
  • Retrieving and displaying metrics snapshots

• examples/metrics_demo.py 📊 NEW - Comprehensive metrics demo

  python examples/metrics_demo.py
  

  Demonstrates:

  • Message throughput and latency tracking
  • Memory cache hit rate monitoring
  • Error tracking and reporting
  • Continuous monitoring patterns

See examples/README_METRICS.md for complete metrics documentation.

Test Files

All test files are located in the tests/ directory:

• tests/test_actions.py - Action system tests
• tests/test_agent_messages.py - Agent messaging tests
• tests/test_async_agent.py - Async functionality tests
• tests/test_advanced_features.py - Advanced features
• tests/test_bindings.py - Basic bindings tests
• tests/test_decorator.py - Action decorator tests
• tests/test_llm_agent.py - LLM agent tests
• tests/test_mesh.py - Mesh operations tests
• tests/test_ollama_simple.py - Ollama connectivity tests
• tests/test_response.py - Response handling tests

API Reference

Core Classes

Agent

Base class for creating custom agents.

class MyAgent(Agent):
    def on_message(self, message: str, context=None) -> str:
        """Handle incoming messages (synchronous)"""
        return "response"

    @Agent.action(name="my_action")
    def custom_action(self, param: str) -> str:
        """Custom action callable by other agents"""
        return f"Processed: {param}"

Methods:

• name() -> str - Get agent name
• send_message(target: str, message: str) - Send message to another agent
• on_message(message: str, context=None) - Override to handle messages

Decorators:

• @Agent.action(name="action_name") - Register a custom action

LlmAgent

LLM-powered agent with fluent builder API.

agent = LlmAgent("name", "ollama::model_name")
agent.with_system_prompt("...")
agent.with_temperature(0.7)
agent.with_max_tokens(100)
agent.build()

Builder Methods:

• with_system_prompt(prompt: str) - Set system prompt
• with_temperature(temp: float) - Set temperature (0.0-1.0)
• with_max_tokens(max: int) - Set max tokens
• build() - Finalize configuration

Message Methods:

• send_message(message: str) -> str - Synchronous LLM call
• send_message_async(message: str) -> Awaitable[str] - Async LLM call ✅

PyLocalMesh

Local in-memory mesh for agent communication.

mesh = PyLocalMesh("mesh_name")
mesh.add_agent(agent)
mesh.send_to("agent_name", "message")

Methods:

• add_agent(agent: Agent) - Register an agent
• send_to(target: str, payload: str) - Send message to agent

PyAction

Custom action definition with schema generation.

from ceylon import PyAction

action = PyAction(
    name="my_action",
    description="Action description",
    schema='{"type": "object", ...}'
)

PyToolInvoker

Execute registered actions.

invoker = agent.tool_invoker
result = invoker.invoke("action_name", '{"param": "value"}')

Async Support

✅ Fully Supported Async Features

1. send_message_async() on LlmAgent

• Fully functional and production-ready
• Supports concurrent execution with asyncio
• Proper error propagation

async def example():
    agent = LlmAgent("agent", "ollama::model")
    agent.build()

    # Concurrent queries
    tasks = [agent.send_message_async(q) for q in queries]
    results = await asyncio.gather(*tasks)

2. Async on_message() handlers ✨ NEW

• Now fully supported with thread-local event loops
• Can use async/await in custom agent message handlers
• Supports async actions as well

class MyAgent(Agent):
    async def on_message(self, message, context=None):
        await asyncio.sleep(0.1)  # Async operations work!
        return f"Processed: {message}"

For detailed async examples, see ASYNC_EXAMPLES.md and ASYNC_STATUS.md.

Documentation

• ASYNC_EXAMPLES.md - Comprehensive async examples guide
• ASYNC_STATUS.md - Current status of async features
• examples/README_METRICS.md - Metrics collection and monitoring guide
• Ceylon Docs - Full framework documentation

Requirements

• Python 3.8+
• Rust toolchain (for building from source)
• Ollama (for LLM examples)

Installing Ollama

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Start Ollama
ollama serve

# Pull a model
ollama pull gemma3:latest

Development

Building from Source

cd bindings/python
cargo build --release
pip install -e .

Running Tests

cd bindings/python
python -m pytest tests/

Or run individual tests:

python tests/test_actions.py
python tests/test_agent_messages.py
python tests/test_llm_agent.py

Architecture

Ceylon uses a mesh architecture where agents communicate through a unified mesh abstraction:

┌─────────────────────────────────────┐
│          Application Code           │
│         (Python/Rust)               │
└──────────────┬──────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│         Agent Mesh (Rust)           │
│  ┌──────┐  ┌──────┐  ┌──────┐      │
│  │Agent1│  │Agent2│  │Agent3│      │
│  └──┬───┘  └──┬───┘  └──┬───┘      │
│     └─────────┴─────────┘          │
│      Message Routing & Delivery    │
└─────────────────────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│   Local (In-Memory) or Distributed  │
│      (Network) Communication        │
└─────────────────────────────────────┘

Key Concepts:

• Agents: Autonomous entities that process messages and execute actions
• Mesh: Communication layer that routes messages between agents
• Actions: Callable functions/tools that agents can invoke
• Messages: Data exchanged between agents

Contributing

Contributions are welcome! Please:

1. Check existing issues or create a new one
2. Fork the repository
3. Create a feature branch
4. Make your changes with tests
5. Submit a pull request

License

See the main Ceylon repository for license information.

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Docs: Ceylon Documentation

Roadmap

• Full async/await support for message handlers
• Additional LLM provider integrations
• Distributed mesh implementation
• Agent lifecycle hooks
• Advanced debugging tools
• Performance monitoring

---

Status: Alpha - API may change

For more information about Ceylon, visit the main repository.