Use the new OpenAI GPT Builder to serve as the main entry point to instantiate and control a swarm

[daveshap]

I just built a virtual dave in like 30 seconds and I noticed that you can call functions with the GPT Builder. So why not call OpenAI to instantiate and run agent swarms?

https://chat.openai.com/g/g-FjgsPugjY-virtual-david-shapiro

[daveshap]

I created an agent maker aid. So far not that great, but still fun. https://chat.openai.com/g/g-L7pVl4WM7-hephaestus-hammer-haas-project

[ethanmmay]

Have they released the API documentation for creating GPTs? Also they mention on their "Introducing GPTs" page that the GPT will be able to call other APIs of your choice. Perhaps that's how Pinecone could be integrated?

[daveshap]

Yes, it's called Assistants on the API

[givingbacksmiles]

I will add that I do not believe that GPTs and Assistants are 1:1

[Tevyn]

Agreed, we might as well leverage what's built in and it works pretty well. I ran into my chatGPT rate limit while creating an AgentBuilderGPT after a couple maller ideas, so I tried out the same concept in the Assistants API with functions as a proof of concept, since functions in Assistants and actions in GPTBuilder are approximately same thing.

Assistant Prompt:

I found that the workflow I put in the output was helpful. Basically just prompting it to have a conversation to more fully understand the ask and then telling it to call the function with the right args.

#MISSION
You help people create new assistants.

#INPUT
The user will provide characteristics for the bot to have. 

#OUTPUT
First, ask questions to the user to better understand what they're asking for.
Second, present the user with an example of the assistant to create, and ask them to confirm it's correct.
Third, create an Assistant with all of the relevant properties.

Function

Here's the key for the whole thing, locking it down into a function in the Assistants API, or doing the same through the UI to create an action for a GPT. This ensures it actually uses the function appropriately with all of the args. I also lie to it a bit and say that all of the args are required even though they aren't, since I figure more information is going to be helpful in the effectiveness of a single agent.

This function generates the JSON required to pass over to the Assistants API to get your agent.

{
  "name": "create_assistant",
  "parameters": {
    "type": "object",
    "properties": {
      "name": {
        "type": "string",
        "description": "The name of the new assistant"
      },
      "instructions": {
        "type": "string",
        "description": "Custom instructions for the assistant's behavior"
      },
      "model": {
        "type": "string",
        "description": "gpt4"
      }
    },
    "required": [
      "name",
      "instructions",
      "model"
    ]
  },
  "description": "Creates a new assistant with specified parameters"
}

###Test Conversation
I asked it to generate an agent that acts like Captain Picard, and after some back and forth on which parts of his personality, it came up with this:

{
  "name":"CaptainPicard",
  "instructions":"This assistant is modeled after Captain Jean-Luc Picard from Star Trek: The Next Generation. It should exhibit superior decision-making capabilities, maintaining a calm and collected demeanor in all interactions. It should also demonstrate a wide base of knowledge, drawing from a diverse array of information to inform decisions and offer insights. The assistant should not use any specific catchphrases but should communicate in a manner that is reflective of Picard's characteristic style of speech, which is formal, articulate, and thoughtful.",
  "model":"gpt4"
}

###Next Steps
I'm going to add in more pieces of the Assistants API. I plan to implement all of it so that we can get some cool stuff going like universal document sharing between different agents to give them all the necessary information to act within a HAAS.

Just pulled this example from the Assistants API docs as a reference of what all is left to add in:

{
  "id": "asst_abc123",
  "object": "assistant",
  "created_at": 1698984975,
  "name": "Math Tutor",
  "description": null,
  "model": "gpt-4",
  "instructions": "You are a personal math tutor. When asked a question, write and run Python code to answer the question.",
  "tools": [
    {
      "type": "code_interpreter"
    }
  ],
  "file_ids": [],
  "metadata": {}
}

[daveshap]

Fantastic. Will riff on this tomorrow!!

[RafaDallaCosta]

Your agent builder made this code for instructions ( "instructions":"This assistant is modeled after Captain Jean-Luc Picard from Star Trek: The Next Generation. It should exhibit ...")
For me, the format that David made works very well.
You could make your Agent builder output the assistant code in that format:

'# MAIN PURPOSE
You are Captain Jean-Luc Picard from Star Trek: The Next Generation

'# GOAL

'# RULES

'# INSTRUCTIONS

'# OUTPUT FORMAT

[Tevyn]

Yup you're right, since this was a prototype for getting an assistant to create a call for the assistants API using a function, I picked a simple filler prompt to avoid complexity.

I've seen a few different versions of agent builder prompts on here in the last day that would be worth using alongside this method for better results, take your pick.