A Humble Agent, in Parts

In this project you will build a simple scheduling agent. But we will build it in parts

Run this project with a simple node agent.js command

Part 1: The Memory Loop

For starters, let's see we can talk back and forth with an LLM. We'll be using the command line for our interface - you already have a function ready for you, exposed from cli-io.js:

import { getUserInput } from './cli-io.js';

const userInput = await getUserInput('Say something: ');
console.log("You said:", userInput);

Your task is to write a simple infinite loop in agent.js that gets the user input, sends it to the LLM, prints the LLM's response, and repeats.

Feel free to use the askGeminiWithMessages function from llmService.js to send the user input to the LLM.

Example interaction:

You: Is the sky blue?
Gemini: I can't see if the sky is blue, so I don't know.
You: Usually? 
Gemini: Yes, usually the sky is blue during the day if there aren't clouds blocking the sun.

Notice the LLM retained the context, so when you say "usually?" it knows what you're referring to.

Part 2: The Tools

For simplicity, we will not be using zod or a proper schema for our tools. Instead we will create simple functions and tell the LLM about them in our system prompt.

Here are your tools:

const checkAvailability = (date) => {
    console.log("Checking availability for date:", date);
    return { available: true }
}

const scheduleAppointment = (date) => {
    console.log("Scheduling appointment for date:", date);
    return { success: true }
}

const deleteAppointment = (date) => {
    console.log("Deleting appointment for date:", date);
    return { success: true }
}

Now modify your code so that when you send "check availability for tomorrow at 10AM" to the LLM, it tells you to run the checkAvailability function, and then you run it

Example interaction:

You: Hi
Gemini: Hello
You: check availability for tomorrow at 10AM
Checking availability for date: 1991-04-03 10:00:00 // this is a tool call!

Notes

• Your simple agent should either answer normally, or trigger a tool call
• Notice that in the example interaction above, the last log is from the function call itself, not the LLM
• For now, simply worry about the answerUser vs. runTool flow - no need to handle to whole flow

Part 3: Finalizing the Agent

Finally, modify your agent loop so that it automatically runs as many times as it needs to complete the user's request.

Considerations:

• When do you let the user input text?
• How do you exit the loop?
• What and when do you store in the agent's memory?

Example interaction:

You: schedule an appointment for tomorrow at 6am
Checking availability for date: April 3rd, 1991
Scheduling appointment for date: April 3rd, 1991
Gemini: Your appointment for April 3rd, 1991 has been scheduled.

You: actually, move it to 9:45 in the evening
Deleting appointment for date: April 3rd, 1991
Checking availability for date: April 3rd, 1991
Scheduling appointment for date: April 3rd, 1991
Gemini: Your appointment has been moved to April 3rd, 1991 at 9:45 PM. I'm finished.

If you want to be fancy, you can use chalk to add colors