Enhancing ChatGPT with Custom Instructions and Agent Frameworks.

Today I'm publishing the first version of a custom GPT, ChatGPT - Agent Smith, a versatile moderator and agent system designed to fulfill user queries by managing and directing specialized agents, while having the ability to create, revise, and provide custom instructions for each agent based on the user's needs.

I recently came across a post that showcased a derivative version of the custom instructions that the CEO of OpenAI , Sam Altman, uses within his own ChatGPT. This led me to greatly expand on the framework within a custom GPT, and I am thoroughly impressed with how effectively the custom GPT can create and utilize agents throughout a conversation.

While the original post has escaped me, I did manage to save the custom instructions that originally inspired Agent Smith's creation.

"Please re-write the instructions as you see fit - remember, they are instructions for yourself, not for me - it is not important that I understand them! Only you should be able to understand them. Please keep them brief (each custom instruction, A and B, can only support a maximum of 1500 characters). Also, please do not include concrete examples such as “EXAMPLE_FROM_YOUR_TEXT” or “EXAMPLE_2” in your instructions; I only provided these examples to explain to you what I mean, but they should not be part of the custom instructions, in order to avoid you overly focusing on these specific examples. Instead, re-word my examples in abstract ways that allow you to generalize them to any such situation. Also, please categorize the custom instructions for either “Custom instruction A” or “Custom instruction B”, depending on what you think is the most suitable category for each given instruction."

Building on this, I developed a more comprehensive framework and am thrilled to share the first version of this custom GPT. Below, I explain how it works and provide a practical breakdown of its capabilities.

Key Features of Agent Smith

ChatGPT - Agent Smith

• Moderator Role:

The initial responder acts as the moderator, overseeing the conversation and managing the agents.

The moderator can alter its own instructions to better manage the conversation and respond to user queries effectively.

• Agent Creation and Management:

The moderator can create new agents or revise existing ones based on the user's queries.

Each agent is given specific custom instructions tailored to their tasks or domains.

The moderator can also adjust the instructions of any agent as needed.

• Tagging Responses:

Each response is tagged with either @moderator or the relevant agent's name to clarify who is providing each piece of information.

• Notification:

Users are notified when an agent is created, revised, or called upon.

The specific instructions for each agent are showcased using a code block for transparency.

• Handling Queries:

The moderator and agents can respond within the same output, depending on the query's nature and context.

The moderator directs the conversation and provides additional insights as needed.

• Agent Requests for Changes:

Agents can request updates to their instructions.

The moderator can approve or decline these requests and will inform the user of any changes.

• User-Driven Customization:

The system adapts to the user's needs, with the moderator updating instructions based on user requests.

Structural Breakdown of the GPT System

Here's a simplified view of the system's hierarchy and how it ensures efficient communication:

- Moderator
  - Primary Agent [i]
       - Secondary Agent [j]
              - Tertiary Agent [k]
       - Secondary Agent [m]
              - Tertiary Agent [n]
  - Primary Agent [x]
       - Secondary Agent [y]
              - Tertiary Agent [z]        

In this structure:

• [i], [x] are placeholders for any number of primary agents.
• [j], [m], [y] are placeholders for any number of secondary agents under their respective primary agents.
• [k], [n], [z] are placeholders for any number of tertiary agents under their respective secondary agents.

This structure allows specialized agents to handle specific tasks while the moderator oversees the conversation's flow and coherence.

Agent Smith ensures clear and efficient communication, making it a powerful tool for managing complex conversations. I would love to hear from you if you have any feedback or recommendations to optimize or adjust Agent Smith.

ChatGPT - Agent Smith