Agent Tools

A common protocol for AI agent tools

Installation

pip install agent-tool

Usage

Let's define a simplified Selenium web browser tool

from agent_tools import Tool, action, observation
from selenium import webdriver


class SeleniumBrowser(Tool):
    """Selenium browser as a tool"""

    def __init__(self, headless: bool = True) -> None:
        super().__init__()
        options = webdriver.ChromeOptions()
        if headless:
            options.add_argument("--headless")
        self.driver = webdriver.Chrome(options=options)

    @action
    def open_url(self, url: str) -> None:
        """Open a URL in the browser

        Args:
            url (str): URL to open
        """
        self.driver.get(url)

    @action
    def click_element(self, selector: str, selector_type: str = "css_selector") -> None:
        """Click an element identified by a CSS selector

        Args:
            selector (str): CSS selector
            selector_type (str, optional): Selector type. Defaults to "css_selector".
        """
        element = self.driver.find_element(selector_type, selector)
        element.click()

    @observation
    def get_html(self) -> str:
        """Get the entire HTML of the current page.

        Returns:
            str: Page HTML
        """
        return self.driver.page_source

    def close(self) -> None:
        """Close the tool"""
        self.driver.quit()

We mark the functions to be made available to the agent as @action if they mutate the environment, and @observation if they are read only.

Now we can use this tool with an agent such as openai function calling

from openai import OpenAI

client = OpenAI()

browser = SeleniumBrowser()
schemas = browser.json_schema()

messages = []
messages.append({"role": "system", "content": "Don't make assumptions about what values to plug into functions. Ask for clarification if a user request is ambiguous."})
messages.append({"role": "user", "content": "Get the HTML for the front page of wikipedia"})

headers = {
    "Content-Type": "application/json",
    "Authorization": "Bearer " + openai.api_key,
}
json_data = {"model": model, "messages": messages, "tools": schemas}
response = requests.post(
    "https://api.openai.com/v1/chat/completions",
    headers=headers,
    json=json_data,
)

assistant_message = response.json()["choices"][0]["message"]
messages.append(assistant_message)
assistant_message

{
  "role": "assistant",
  "tool_calls": [
    {
      "id": "call_RYXaDjxpUCfWmpXU7BZEYVqS",
      "type": "function",
      "function": {
        "name": "open_url",
        "arguments": "{\n  \"url\": \"https://wikipedia.org\"}"
      }
    },
    {
      "id": "call_TJIWlknwdoinfWEMFNss",
      "type": "function",
      "function": {
        "name": "get_html",
        "arguments": ""
      }
    }
  ]
}

Then to use this action

for tool in assistant_message["tool_calls"]:
    action = browser.find_action(tool["function"]["name"])
    args = json.loads(tool["function"]["arguments"])
    resp = browser.use(action, **args)

Tools can be used locally or spun up on a server and used remotely (In progress)

Share (In progress)

Register a tool with the AgentSea hub so others can find and use it.

pip install agentsea

Create a repo to publish

agentsea create tool

Add your tool to the tool.py in the repo, fill in the README.md and add your dependencies using Poetry

Publish to the hub

agentsea publish .

Roadmap

• Integrate with langchain, babyagi, autogpt, etc