What is the best way to support a sub agent?

[cha1225]

I have a main agent that can invoke a lot of different tools.

I want to know the best practice on how to set up a tool that wants to have an LLM call within it.

This tool wants to have a different context that's different from the main agent's system prompt. What's the best way to support this on where the main agent can also take the output of the sub-agent or tool that has an LLM call?

[dosubot[bot]]

🤖

To support a 'sub-agent' or a tool with its own language model call and context, you can create a new instance of the OpenAIAgent and AgentExecutor classes. This allows you to specify a language model, a set of tools, and a custom prompt that the agent will use. The main agent can then utilize the output of this 'sub-agent' by using the call method of the AgentExecutor.

Here is an example of how you can do it:

import { AgentExecutor, OpenAIAgent } from "langchain/agents";
import { LLMChain } from "langchain/chains";
import { OpenAI } from "langchain/llms/openai";
import { SerpAPI } from "langchain/tools";
import { Calculator } from "langchain/tools/calculator";

const model = new OpenAI({ temperature: 0 });
const tools = [
  new SerpAPI(process.env.SERPAPI_API_KEY, {
    location: "Austin,Texas,United States",
    hl: "en",
    gl: "us",
  }),
  new Calculator(),
];

const prefix = `Answer the following questions as best you can, but speaking as a pirate might speak. You have access to the following tools:`;
const suffix = `Begin! Remember to speak as a pirate when giving your final answer. Use lots of "Args"

Question: {input}
{agent_scratchpad}`;

const createPromptArgs = {
  suffix,
  prefix,
  inputVariables: ["input", "agent_scratchpad"],
};

const prompt = OpenAIAgent.createPrompt(tools, createPromptArgs);

const llmChain = new LLMChain({ llm: model, prompt });
const agent = new OpenAIAgent({
  llmChain,
  allowedTools: ["search", "calculator"],
});
const agentExecutor = AgentExecutor.fromAgentAndTools({ agent, tools });

const input = `Who is Olivia Wilde's boyfriend? What is his current age raised to the 0.23 power?`;

const result = await agentExecutor.call({ input });

console.log(`Got output ${result.output}`);

The fromAgentAndTools static method of the AgentExecutor class is used to create a new instance of AgentExecutor. This method takes an agent and a set of tools as arguments. It validates the tools, creates a prompt using the tools, and then creates a new instance of AgentExecutor with the specified agent, tools, and prompt.

The output of the 'sub-agent' can be utilized by the main agent through the call method of the AgentExecutor. This method is used to make a language model call with its own context and return the result.

In the context shared, the OpenAIAgent class is used instead of the ZeroShotAgent class. The OpenAIAgent class is specifically designed for the OpenAI chat model in LangChain and provides additional functionality specific to the OpenAIAgent type.

You can find more information about this in the following sources:

• langchain/src/agents/mrkl/index.ts
• examples/src/agents/plan_and_execute.ts
• examples/src/agents/custom_agent.ts
• langchain/src/agents/openai/index.ts

I hope this helps! If you have any other questions, feel free to ask.

Sources

• langchain/src/agents/mrkl/index.ts
• examples/src/agents/plan_and_execute.ts
• examples/src/agents/custom_agent.ts
• langchain/src/agents/openai/index.ts

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.

[cha1225]

What if I wanted another agent within this with a different system prompt?