| Name | panel-of-agents JSON |
| Version |
0.1.0
JSON |
| download |
| home_page | None |
| Summary | A framework for orchestrating multiple AI agents in collaborative discussions and tasks |
| upload_time | 2025-02-04 02:58:04 |
| maintainer | None |
| docs_url | None |
| author | None |
| requires_python | >=3.10 |
| license | None |
| keywords |
agents
ai
automation
llms
|
| VCS |
 |
| bugtrack_url |
|
| requirements |
No requirements were recorded.
|
| Travis-CI |
No Travis.
|
| coveralls test coverage |
No coveralls.
|
# PanelOfAgents
A flexible and powerful framework for orchestrating multiple AI agents in domain-specific applications, with support for any LLM backend.
## Introduction
This framework enables seamless orchestration of multiple AI agents for domain-specific applications. Inspired by OpenAI's swarm framework but addressing its limitations, our solution offers greater flexibility in agent interactions without being tied to specific LLM implementations.
### Purpose
- Create a simple yet powerful multi-agent orchestration framework
- Enable flexible task handoffs between agents in single-domain applications
- Provide a production-ready alternative to experimental frameworks
- Support integration with various LLM implementations
### Why This Framework?
Unlike OpenAI's swarm framework, which is limited to GPT models and requires predefined handoffs, this framework offers:
- Model-agnostic implementation
- Dynamic agent interactions
- Production-ready architecture
- Flexible integration options
## Features
- **🚀 Flexible Multi-agent Orchestration** - Dynamic task distribution and collaboration between agents
- **⚡️ Real-time Communication** - Token-by-token streaming with minimal latency
- **🔄 Rich Context Sharing** - Comprehensive context including conversation history, action results, and artifacts
- **🔌 Universal LLM Support** - Compatible with any LangChain BaseChatModel implementation
- **🎯 Targeted Agent Usage** - Smart agent selection based on capabilities rather than predefined rules
- **⚙️ Developer-First Design** - Flexible state management and easy service integration
## Installation
```bash
pip install panel-of-agents
```
## Usage
### Context
The `Context` object is the central object in the framework. It contains the state of the conversation, the artifacts, and any custom properties. It is shared between all agents.
The `Context` can be initialized with the following parameters:
- `current_question`: The current question to be answered.
- `conversation_history`: A list of messages between the user and the agents. This is a list of `HumanMessage` and `AIMessage` objects from LangChain.
- `artifacts`: A dictionary of artifacts. The key is the index of the conversation turn and the value is an `Artifact` object.
- `target_agent` (optional): The agent that will respond to the user.
- `custom_props` (optional): A dictionary of custom properties. The key is the name of the property and the value is a list of `CustomProp` objects.
#### Initializing an empty context
```python
from panel_of_agents.context import Context
context = Context("Hi", [], {})
```
#### Initializing a context with a conversation history
```python
from langchain_core.messages import HumanMessage, AIMessage
context = Context("Hi", [HumanMessage("Hi"), AIMessage("Hi, how can I help you today?")], {})
```
### Decorators
`panel_of_agents` uses decorators to indicate the actions/tools that an agent can use.
#### `capability`
The `capability` decorator is used to indicate that a function is a capability.
```python
from panel_of_agents.decorators import capability
class SomeAgent(Agent):
...
@capability
def my_capability(self, context: Context) -> CapabilityResult:
return CapabilityResult(result="My capability was successful", artifact=None)
```
Capabilities which are the actions that an agent can perform must return a `CapabilityResult` object.
The `CapabilityResult` object contains the following parameters:
- `result`: The result of the capability. The output of the function.
- `artifact`: An optional artifact.
#### Artifacts
Artifacts are outputs of capabilities that may be stored and persisted across multiple conversation turns.
```python
from panel_of_agents.types.context import Artifact
artifact = Artifact(author="SomeAgent", data={"some_data": "some_data"}, artifact_type=ArtifactType.INTERNAL)
```
The `artifact_type` can be either `INTERNAL` or `USER`.
- `INTERNAL`: The artifact is internal to the framework and is not visible to the user.
- `USER`: The artifact is visible to the user and can be used by other capabilities.
Capabilities which create artifacts must be decorated with the `creates_artifact` decorator.
```python
from panel_of_agents.decorators import capability, creates_artifact
class SomeAgent(Agent):
...
@creates_artifact("A list of users")
@capability
def my_capability(self, context: Context) -> CapabilityResult:
artifact = Artifact(author=self.name, data={"some_data": "some_data"}, artifact_type=ArtifactType.INTERNAL)
return CapabilityResult(result="My capability was successful", artifact=artifact)
```
### Agents
An agent is a self-contained entity that can perform various related actions. In `panel_of_agents`, custom agents can be created by extending the `Agent` class.
An agent requires the following parameters:
- `name`: The name of the agent.
- `personal_biography`: A short biography of the agent.
- `public_biography`: A short biography of the agent that is visible to other agents.
- `model`: The model to use for the agent.
- `max_tries` (optional): The maximum number of tries for an agent to perform an action. Defaults to 3.
#### Creating an agent
```python
class AdditionAgent(Agent):
def __init__(self, model: BaseChatModel, max_tries: int = 3):
personal_biography = "You are an agent that performs addition on two numbers."
public_biography = "An agent that performs addition on two numbers."
super().__init__(
name="Addition Agent",
personal_biograhpy=personal_biography,
public_biograhpy=public_biography,
model=model,
max_tries=max_tries
)
@agent_capability
def add(self, a: float, b: float) -> CapabilityResult:
result = a + b
return CapabilityResult(result=result, artifact=None)
addition_agent = AdditionAgent(model, max_tries=5)
```
### Moderator
The `Moderator` is responsible for orchestrating the conversation between the agents. It is responsible for selecting the initial agent (when `target_agent` is not set), and for controlling the number of times agents execute to answer a given question.
#### Creating a moderator
The `Moderator` requires the following parameters:
- `panel_of_agents`: A list of agents to be used in the conversation. These are instances of the subclassed `Agent` class.
- `leader`: The agent that will respond to the user.
- `moderator_cast_vote`: Whether the moderator will cast a vote on whether an agent is capable of answering a given question. If not set, each agent will vote on whether they are capable of answering the question.
```python
from panel_of_agents.moderator import Moderator
moderator = Moderator(
panel_of_agents=[AdditionAgent(), SubtractionAgent()],
leader="Addition Agent",
moderator_cast_vote=True
)
```
### Transmitter
The `Transmitter` is responsible for transmitting the conversation between the agents. It is responsible for formatting the conversation for the user and for streaming the conversation to the user.
#### Creating a transmitter
```python
from panel_of_agents.transmitter import Transmitter
transmitter = Transmitter(moderator)
```
#### Invoking the transmitter
##### Streaming
```python
for token in transmitter.invoke_moderator(context, stream=True):
print(token, end="", flush=True)
```
##### Non-streaming
```python
response = transmitter.invoke_moderator(context, stream=False)
```
#### Raw Feed
The `panel_of_agents` framework utilizes a custom communication protocol, for consistent performance, output containing the raw response should be used as the conversation history.
The `raw_feed` can be accessed from the `transmitter` object after an invocation is complete.
```python
response = transmitter.invoke_moderator(context, stream=False)
conversation_history.append(AIMessage(content=transmitter.raw_feed))
```
Raw data
{
"_id": null,
"home_page": null,
"name": "panel-of-agents",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.10",
"maintainer_email": "Mohammed Ali <splitwireml@gmail.com>",
"keywords": "agents, ai, automation, llms",
"author": null,
"author_email": "Mohammed Ali <splitwireml@gmail.com>",
"download_url": "https://files.pythonhosted.org/packages/f6/da/9421b57513c0211d4818270015e6d1dd890c5467c9c051dd0ae8a60e9766/panel_of_agents-0.1.0.tar.gz",
"platform": null,
"description": "# PanelOfAgents\n\nA flexible and powerful framework for orchestrating multiple AI agents in domain-specific applications, with support for any LLM backend.\n\n## Introduction\n\nThis framework enables seamless orchestration of multiple AI agents for domain-specific applications. Inspired by OpenAI's swarm framework but addressing its limitations, our solution offers greater flexibility in agent interactions without being tied to specific LLM implementations.\n\n### Purpose\n\n- Create a simple yet powerful multi-agent orchestration framework\n- Enable flexible task handoffs between agents in single-domain applications\n- Provide a production-ready alternative to experimental frameworks\n- Support integration with various LLM implementations\n\n### Why This Framework?\n\nUnlike OpenAI's swarm framework, which is limited to GPT models and requires predefined handoffs, this framework offers:\n- Model-agnostic implementation\n- Dynamic agent interactions\n- Production-ready architecture\n- Flexible integration options\n\n## Features\n\n- **\ud83d\ude80 Flexible Multi-agent Orchestration** - Dynamic task distribution and collaboration between agents\n- **\u26a1\ufe0f Real-time Communication** - Token-by-token streaming with minimal latency\n- **\ud83d\udd04 Rich Context Sharing** - Comprehensive context including conversation history, action results, and artifacts\n- **\ud83d\udd0c Universal LLM Support** - Compatible with any LangChain BaseChatModel implementation\n- **\ud83c\udfaf Targeted Agent Usage** - Smart agent selection based on capabilities rather than predefined rules\n- **\u2699\ufe0f Developer-First Design** - Flexible state management and easy service integration\n\n## Installation\n\n```bash\npip install panel-of-agents\n```\n\n## Usage\n\n### Context\n\nThe `Context` object is the central object in the framework. It contains the state of the conversation, the artifacts, and any custom properties. It is shared between all agents.\n\nThe `Context` can be initialized with the following parameters:\n- `current_question`: The current question to be answered.\n- `conversation_history`: A list of messages between the user and the agents. This is a list of `HumanMessage` and `AIMessage` objects from LangChain.\n- `artifacts`: A dictionary of artifacts. The key is the index of the conversation turn and the value is an `Artifact` object.\n- `target_agent` (optional): The agent that will respond to the user.\n- `custom_props` (optional): A dictionary of custom properties. The key is the name of the property and the value is a list of `CustomProp` objects.\n\n#### Initializing an empty context\n\n```python\nfrom panel_of_agents.context import Context\n\ncontext = Context(\"Hi\", [], {})\n```\n\n#### Initializing a context with a conversation history\n\n```python\nfrom langchain_core.messages import HumanMessage, AIMessage\n\ncontext = Context(\"Hi\", [HumanMessage(\"Hi\"), AIMessage(\"Hi, how can I help you today?\")], {})\n```\n\n### Decorators\n\n`panel_of_agents` uses decorators to indicate the actions/tools that an agent can use.\n\n#### `capability`\n\nThe `capability` decorator is used to indicate that a function is a capability.\n\n```python\nfrom panel_of_agents.decorators import capability\n\nclass SomeAgent(Agent):\n ...\n\n @capability\n def my_capability(self, context: Context) -> CapabilityResult:\n return CapabilityResult(result=\"My capability was successful\", artifact=None)\n```\n\nCapabilities which are the actions that an agent can perform must return a `CapabilityResult` object.\n\nThe `CapabilityResult` object contains the following parameters:\n- `result`: The result of the capability. The output of the function.\n- `artifact`: An optional artifact.\n\n#### Artifacts\n\nArtifacts are outputs of capabilities that may be stored and persisted across multiple conversation turns.\n\n```python\nfrom panel_of_agents.types.context import Artifact\n\nartifact = Artifact(author=\"SomeAgent\", data={\"some_data\": \"some_data\"}, artifact_type=ArtifactType.INTERNAL)\n```\n\nThe `artifact_type` can be either `INTERNAL` or `USER`.\n\n- `INTERNAL`: The artifact is internal to the framework and is not visible to the user.\n- `USER`: The artifact is visible to the user and can be used by other capabilities.\n\nCapabilities which create artifacts must be decorated with the `creates_artifact` decorator.\n\n```python\nfrom panel_of_agents.decorators import capability, creates_artifact\n\nclass SomeAgent(Agent):\n ...\n\n @creates_artifact(\"A list of users\")\n @capability\n def my_capability(self, context: Context) -> CapabilityResult:\n artifact = Artifact(author=self.name, data={\"some_data\": \"some_data\"}, artifact_type=ArtifactType.INTERNAL)\n return CapabilityResult(result=\"My capability was successful\", artifact=artifact)\n```\n\n### Agents\n\nAn agent is a self-contained entity that can perform various related actions. In `panel_of_agents`, custom agents can be created by extending the `Agent` class.\n\nAn agent requires the following parameters:\n- `name`: The name of the agent.\n- `personal_biography`: A short biography of the agent.\n- `public_biography`: A short biography of the agent that is visible to other agents.\n- `model`: The model to use for the agent.\n- `max_tries` (optional): The maximum number of tries for an agent to perform an action. Defaults to 3.\n\n#### Creating an agent\n\n```python\n\nclass AdditionAgent(Agent):\n def __init__(self, model: BaseChatModel, max_tries: int = 3):\n personal_biography = \"You are an agent that performs addition on two numbers.\"\n public_biography = \"An agent that performs addition on two numbers.\"\n\n super().__init__(\n name=\"Addition Agent\",\n personal_biograhpy=personal_biography,\n public_biograhpy=public_biography,\n model=model,\n max_tries=max_tries\n )\n\n @agent_capability\n def add(self, a: float, b: float) -> CapabilityResult:\n result = a + b\n return CapabilityResult(result=result, artifact=None)\n\n\naddition_agent = AdditionAgent(model, max_tries=5)\n```\n\n### Moderator\n\nThe `Moderator` is responsible for orchestrating the conversation between the agents. It is responsible for selecting the initial agent (when `target_agent` is not set), and for controlling the number of times agents execute to answer a given question.\n\n#### Creating a moderator\n\nThe `Moderator` requires the following parameters:\n- `panel_of_agents`: A list of agents to be used in the conversation. These are instances of the subclassed `Agent` class.\n- `leader`: The agent that will respond to the user.\n- `moderator_cast_vote`: Whether the moderator will cast a vote on whether an agent is capable of answering a given question. If not set, each agent will vote on whether they are capable of answering the question.\n\n```python\nfrom panel_of_agents.moderator import Moderator\n\nmoderator = Moderator(\n panel_of_agents=[AdditionAgent(), SubtractionAgent()],\n leader=\"Addition Agent\",\n moderator_cast_vote=True\n)\n```\n\n### Transmitter\n\nThe `Transmitter` is responsible for transmitting the conversation between the agents. It is responsible for formatting the conversation for the user and for streaming the conversation to the user.\n\n#### Creating a transmitter\n\n```python\nfrom panel_of_agents.transmitter import Transmitter\n\ntransmitter = Transmitter(moderator)\n```\n\n#### Invoking the transmitter\n\n##### Streaming\n```python\n\nfor token in transmitter.invoke_moderator(context, stream=True):\n print(token, end=\"\", flush=True)\n```\n\n##### Non-streaming\n\n```python\nresponse = transmitter.invoke_moderator(context, stream=False)\n```\n\n#### Raw Feed\n\nThe `panel_of_agents` framework utilizes a custom communication protocol, for consistent performance, output containing the raw response should be used as the conversation history.\n\nThe `raw_feed` can be accessed from the `transmitter` object after an invocation is complete.\n\n```python\nresponse = transmitter.invoke_moderator(context, stream=False)\n\nconversation_history.append(AIMessage(content=transmitter.raw_feed))\n```\n\n\n\n",
"bugtrack_url": null,
"license": null,
"summary": "A framework for orchestrating multiple AI agents in collaborative discussions and tasks",
"version": "0.1.0",
"project_urls": {
"Homepage": "https://github.com/LighthouseAI-saas/PanelOfAgents",
"Repository": "https://github.com/LighthouseAI-saas/PanelOfAgents"
},
"split_keywords": [
"agents",
" ai",
" automation",
" llms"
],
"urls": [
{
"comment_text": null,
"digests": {
"blake2b_256": "f76c81af02df039da350ad91f26f5c9d2fe99add0c086ef73317227457dd02f0",
"md5": "0bb2377aae25a9bef308e6ba7693c8a1",
"sha256": "1120f80f6915acbcb5aa0c75a8ef7a34fd15032f9167c739a38767404097b7d3"
},
"downloads": -1,
"filename": "panel_of_agents-0.1.0-py3-none-any.whl",
"has_sig": false,
"md5_digest": "0bb2377aae25a9bef308e6ba7693c8a1",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.10",
"size": 38103,
"upload_time": "2025-02-04T02:58:02",
"upload_time_iso_8601": "2025-02-04T02:58:02.124744Z",
"url": "https://files.pythonhosted.org/packages/f7/6c/81af02df039da350ad91f26f5c9d2fe99add0c086ef73317227457dd02f0/panel_of_agents-0.1.0-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": null,
"digests": {
"blake2b_256": "f6da9421b57513c0211d4818270015e6d1dd890c5467c9c051dd0ae8a60e9766",
"md5": "469f37fbec1273d1406903a56fa9bde0",
"sha256": "bc293c01fc080ec9ad84719c0b598884c18ac49cca0c846e31c6cddda40e9948"
},
"downloads": -1,
"filename": "panel_of_agents-0.1.0.tar.gz",
"has_sig": false,
"md5_digest": "469f37fbec1273d1406903a56fa9bde0",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.10",
"size": 72659,
"upload_time": "2025-02-04T02:58:04",
"upload_time_iso_8601": "2025-02-04T02:58:04.190856Z",
"url": "https://files.pythonhosted.org/packages/f6/da/9421b57513c0211d4818270015e6d1dd890c5467c9c051dd0ae8a60e9766/panel_of_agents-0.1.0.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2025-02-04 02:58:04",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "LighthouseAI-saas",
"github_project": "PanelOfAgents",
"github_not_found": true,
"lcname": "panel-of-agents"
}
