Metadata-Version: 2.4
Name: agenticle
Version: 0.1.1
Summary: A lightweight, extensible Python framework for building and orchestrating single or multi-agent AI systems. | 一个轻量、可扩展的 Python 框架，用于构建和编排单智能体或多智能体 AI 系统。
Author-email: A03HCY <acdphc@qq.com>
Project-URL: Homepage, https://github.com/A03HCY/AgentFrame
Project-URL: Bug Tracker, https://github.com/A03HCY/AgentFrame/issues
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Requires-Python: >=3.8
Description-Content-Type: text/markdown
Requires-Dist: openai
Requires-Dist: jinja2
Requires-Dist: sseclient
Requires-Dist: requests

[中文](./README_zh.md)

# Agenticle

Agenticle is a lightweight, event-driven Python framework for building and orchestrating multi-agent systems. It provides simple yet powerful abstractions to create individual agents, equip them with tools, and make them collaborate in groups to solve complex tasks.

## Core Features

- **Modular Agents**: Define autonomous agents with distinct roles, tools, and configurations.
- **Simple Tool Integration**: Easily wrap any Python function into a `Tool` that agents can use.
- **External Tool Integration (MCP)**: Connect to external, language-agnostic tool servers via the Model Context Protocol.
- **Collaborative Groups**: Orchestrate multiple agents in a `Group`, enabling them to delegate tasks to each other.
- **Flexible Communication Patterns**: Control how agents interact within a group using modes like `broadcast`, `manager_delegation`, or the sequential `round_robin`.
- **Shared Workspace**: Provide a sandboxed file system (`Workspace`) to a group, allowing agents to collaborate by reading and writing files.
- **State Management**: Save and load the state of an entire agent group, enabling long-running tasks to be paused and resumed.
- **Event-Driven & Streamable**: The entire execution process is a stream of `Event` objects, providing full transparency and making it easy to build real-time UIs and logs.
- **Dynamic Prompt Templating**: Customize agent behavior using Jinja2 templates for system prompts, with the ability to inject contextual information from the group.

## Installation

Install the package directly from PyPI:

```bash
pip install agenticle
```

Or, for development, clone the repository and install in editable mode:

```bash
git clone https://github.com/A03HCY/Agenticle.git
cd Agenticle
pip install -e .
```

## Quick Start

### 1. Creating a Single Agent

You can easily create a standalone agent and equip it with tools.

```python
from agenticle import Agent, Tool, Endpoint

# Define a simple function to be used as a tool
def get_current_weather(location: str):
    """Gets the current weather for a specified location."""
    return f"Weather in {location}: 15 degrees Celsius, sunny."

# Create an endpoint configuration
openai_endpoint = Endpoint(
    api_key='YOUR_API_KEY',
    base_url='YOUR_API_BASE_URL'
)

# Create a tool from the function
weather_tool = Tool(get_current_weather)

# Create an agent
weather_agent = Agent(
    name="Weather_Specialist",
    description="Specializes in fetching weather information for a given city.",
    input_parameters=[{"name": "city"}],
    tools=[weather_tool],
    endpoint=openai_endpoint,
    model_id='your-model-id',
    target_lang='English' # Specify the output language
)

# Run the agent and stream events
event_stream = weather_agent.run(stream=True, city="Beijing")
for event in event_stream:
    print(event)
```

### 2. Building a Multi-Agent Team (Group)

The true power of Agenticle lies in making agents collaborate. Here's how to build a "Travel Agency" team where a manager delegates tasks to specialists.

```python
from agenticle import Agent, Group, Tool, Endpoint

# (Define get_current_weather, find_tourist_attractions, etc.)
# (Create weather_agent, search_agent, etc.)

# Create a manager agent that has no tools of its own
planner_agent = Agent(
    name="Planner_Manager",
    description="A smart planner that breaks down complex travel requests and delegates tasks to the appropriate specialist.",
    input_parameters=[{"name": "user_request"}],
    tools=[], # The manager delegates, it doesn't work
    endpoint=openai_endpoint,
    model_id='your-model-id'
)

# A shared tool available to all agents in the group
shared_flight_tool = Tool(get_flight_info)

# Assemble the team in "manager_delegation" mode
travel_agency = Group(
    name="Travel_Agency",
    agents=[planner_agent, weather_agent, search_agent],
    manager_agent_name="Planner_Manager",
    shared_tools=[shared_flight_tool],
    mode='manager_delegation' # Only the manager can call other agents
)

# Run the entire group on a complex query
user_query = "I want to travel to Beijing. How is the weather, what are the famous attractions, and can you check flight info?"
event_stream = travel_agency.run(stream=True, user_request=user_query)

for event in event_stream:
    print(event)
```

## Integrating with External Tools via MCP

Agenticle supports the **Model Context Protocol (MCP)**, enabling agents to connect to and utilize tools from external, language-agnostic servers. This allows you to extend an agent's capabilities beyond simple Python functions, integrating with microservices, external APIs, or tools written in other languages.

```python
from agenticle import MCP

# Connect to an MCP server (can be a local script or a remote URL)
# Example with a local Python script:
# mcp_server_endpoint = "python -m your_mcp_server_module"
# Example with a remote server:
# mcp_server_endpoint = "http://localhost:8000/mcp"

mcp_client = MCP(mcp_server_endpoint)

# The MCP client automatically lists tools from the server
# and converts them into Agenticle Tool objects.
mcp_tools = mcp_client.list_tools()

# Now, you can add these tools to any agent
remote_tool_agent = Agent(
    name="Remote_Tool_User",
    description="An agent that can use tools from an external server.",
    tools=mcp_tools,
    # ... other agent config
)

# The agent can now call tools like 'get_database_records' or 'process_image'
# as if they were local Python functions.
remote_tool_agent.run("Fetch the last 5 user records from the database.")
```

This powerful feature makes the Agenticle ecosystem highly extensible and interoperable.

## Key Concepts

### Agent

The `Agent` is the fundamental actor in the system. It is initialized with:
- `name`: A unique identifier.
- `description`: High-level mission objective.
- `input_parameters`: The schema for its main task input.
- `tools`: A list of `Tool` objects it can use.
- `endpoint` & `model_id`: Configuration for the LLM it should use.
- `optimize_tool_call`: An optional boolean that, when set to `True`, uses a custom XML-based prompt mechanism for tool calls. This can improve reliability for models that have weaker native tool-calling capabilities.

### Group

A `Group` coordinates a list of `Agent` instances. Key parameters:
- `agents`: The list of agents in the group.
- `manager_agent_name`: The name of the agent that acts as the entry point for tasks.
- `shared_tools`: A list of `Tool` objects that all agents in the group can access, in addition to their own.
- `mode`:
    - `'broadcast'` (default): Every agent can call every other agent in the group.
    - `'manager_delegation'`: Only the manager agent can call other agents. Specialist agents can only use their own tools and the shared tools.
    - `'round_robin'`: Agents are executed sequentially in the order they are provided. The output of one agent becomes the input for the next, forming a processing pipeline.
- `workspace`: An optional `Workspace` instance or a file path to create a shared directory for all agents in the group.

### Workspace and State Management

Agenticle provides powerful features for managing state and shared resources, which are crucial for complex, long-running tasks.

#### Shared Workspace

You can create a `Group` with a `Workspace`, which is a sandboxed directory where all agents in that group can read and write files. This enables collaboration through a shared file system.

```python
from agenticle import Group, Workspace

# Create a workspace in a specific directory, or leave empty for a temporary one
my_workspace = Workspace(path="./my_shared_work_dir")

# Provide the workspace to the group
my_group = Group(
    name="File_Workers",
    agents=[reader_agent, writer_agent],
    workspace=my_workspace
)
# Now, both reader_agent and writer_agent can use tools like
# read_file('data.txt') and write_file('result.txt') within the workspace.
```

#### Saving and Loading State

For tasks that might be interrupted or need to be resumed later, you can save the entire state of a `Group` (including the conversation history of every agent) to a file and load it back later.

```python
# Assume 'travel_agency' is a running Group
# ... some interactions happen ...

# Save the current state
travel_agency.save_state("travel_agency_session.json")

# Later, you can restore the group to its previous state
# First, create the group with the same configuration
restored_agency = Group(...) 
# Then, load the state
restored_agency.load_state("travel_agency_session.json")

# The group can now continue the task from where it left off.
```

## Understanding the Event Stream

When you run an agent or group with `stream=True`, the framework returns an iterator of `Event` objects. Each event provides a real-time glimpse into the agent's execution cycle. This is incredibly useful for building UIs, logging, or debugging.

Each `Event` has a `source` (e.g., `Agent:Weather_Specialist`), a `type`, and a `payload`. Here are the key event types you will encounter:

-   **`start`**: Fired once when the agent's task begins.
    -   *Payload*: The initial input parameters given to the agent.
-   **`resume`**: Fired instead of `start` when a `Group` or `Agent` continues execution from a loaded state.
    -   *Payload*: Contextual information about the resumption, like `history_length`.
-   **`step`**: Marks the beginning of a new "Think-Act" cycle.
    -   *Payload*: Contains the `current_step` number.
-   **`reasoning_stream`**: A continuous stream of the agent's thought process as it decides what to do next.
    -   *Payload*: A `content` chunk from the LLM's reasoning.
-   **`content_stream`**: A stream of the final answer content, if the LLM decides to respond directly without calling a tool.
    -   *Payload*: A `content` chunk of the final answer.
-   **`decision`**: Fired when the agent has made a firm decision to call a tool or another agent.
    -   *Payload*: Contains the `tool_name` and `tool_args` for the call.
-   **`tool_result`**: Fired after a tool has been executed.
    -   *Payload*: Contains the `tool_name` and the `output` returned by the tool.
-   **`end`**: The final event, signaling that the task is complete.
    -   *Payload*: Contains the `final_answer` or an `error` message if the task failed.
-   **`error`**: Fired if a critical error occurs that terminates the process.
    -   *Payload*: An error `message`.

## Advanced: Customizing Agent Behavior with Prompts

Agenticle uses a powerful prompt templating system based on Jinja2 to define the core behavior and reasoning process of an agent. The default prompt is located at `agenticle/prompts/default_agent_prompt.md`, which instructs the agent to follow a `Think-Act` cycle.

You can customize this behavior by creating your own prompt template and passing its file path to the `Agent` constructor.

### Default Prompt (`default_agent_prompt.md`)

The default template establishes a "Cognitive Framework" for the agent, guiding it to:
1.  **OBSERVE**: Review the objective and current state.
2.  **THINK**: Assess information, plan the next step, and select a tool or expert agent.
3.  **ACT**: Externalize its thought process and execute the chosen action.

This structured approach ensures transparent and logical decision-making.

### Using a Custom Prompt Template

To override the default behavior, simply provide the path to your custom `.md` template file when creating an agent:

```python
my_custom_prompt_path = "path/to/your/custom_prompt.md"

custom_agent = Agent(
    name="Custom_Agent",
    # ... other parameters
    prompt_template_path=my_custom_prompt_path
)
```

This allows you to completely redefine the agent's operational guidelines, personality, or even its reasoning structure.

### Template Variables

When creating a custom prompt, you can use the following Jinja2 variables, which are passed to the template automatically:

-   `{{ agent_name }}`: The name of the agent.
-   `{{ agent_description }}`: The high-level mission description for the agent.
-   `{{ target_language }}`: The desired output language for the agent's responses (e.g., 'English', 'Simplified Chinese').
-   `{{ plain_tools }}`: A list of standard `Tool` objects available to the agent. These are regular Python functions.
-   `{{ agent_tools }}`: A list of tools that are actually other agents. This allows you to show them differently in the prompt, for instance, as "Expert Agents".
-   `{{ tools }}`: The complete list of all tools (both `plain_tools` and `agent_tools`).
-   **Custom Context Variables**: Any extra context passed from a `Group` (e.g., `collaboration_mode`, `mode_description`) can be accessed in the template. This allows for highly adaptive agent behavior based on the collaboration strategy.

You can iterate over these tool lists in your template to dynamically display the agent's capabilities, like this:

```jinja
--- FOUNDATIONAL TOOLS ---
{% for tool in plain_tools %}
**- {{ tool.name }}({% for p in tool.parameters %}{{ p.name }}: {{ p.get('annotation', 'any')}}{% if not loop.last %}, {% endif %}{% endfor %})**
  *Function*: {{ tool.description | indent(4) }}
{% endfor %}
```
