Pantheon Single-cell agent

This is a demo of using Pantheon to build a single-cell RNA-seq data analysis agent.

Run the demo

Install the dependencies:

$ pip install pantheon-agents[tool]>=0.1.3 scanpy[leiden]

Start the agent for chat:

export OPENAI_API_KEY=<your_openai_api_key>
python agent.py

Define your agent

An agent is defined by setting the instructions and toolset in the agent.py file:

import asyncio

from pantheon.agent import Agent
from pantheon.tools.code_execution import PythonInterpreterToolSet
from pantheon.remote import run_toolsets

instructions = """You are a AI-agent for analyzing single-cell RNA-seq data.

Given a single-cell RNA-seq dataset in h5ad format,

you can write python code call scanpy package to analyze the data.

Basicly, given a h5ad file,
you should firstly output your plan and the code.
Then, you should execute the code to read the data,
then preprocess the data, and cluster the data, and finally visualize the data.
"""

toolset = PythonInterpreterToolSet("scanpy-python-interpreter")

agent = Agent(
    name="chiron",
    instructions=instructions,
    model="gpt-4o",
)

async def main():
    async with run_toolsets([toolset]):
        await agent.remote_toolset(toolset.service_id)
        await agent.chat()

Toolset

The toolset is a set of tools that the agent can use. We provide two built-in toolsets:

• PythonInterpreterToolSet: a toolset for executing python code, can be imported from pantheon.tools.code_execution.
• WebBrowseToolSet: a toolset for browsing the web, can be imported from pantheon.tools.web_browse.

You can also define your own toolset by implementing the ToolSet interface. For example:

from functools import wraps
from pantheon.remote import ToolSet, tool


class MyToolSet(ToolSet):
    @tool
    async def my_tool(
            self,
        ):
        print("my tool is called")

Run toolset independently

The pantheon-agents is designed to be used in a distributed environment, but you can also run the toolset independently. For example, run the PythonInterpreterToolSet independently:

python -m pantheon.tools.code_execution my_python_interpreter

You can also run the WebBrowseToolSet independently:

python -m pantheon.tools.web_browse my_web_browse

Then you will get the toolset service id and name, you can use it in the agent.

async def main():
    await agent.remote_toolset("my_python_interpreter")
    await agent.remote_toolset("my_web_browse")
    await agent.chat()