Processing Agents v1: Setup LLM Utils, setup Modular Agent Base, Initial Prompt Files, Self-debate System

[markokraemer]

Explanation for Implementation

Each agent corresponds to one internal processing layer:

• Planning Agent: Identifies relevant files for editing and creates a detailed task list with exact Instructions for the individual files
• Control Agent: Selects the task to work on, decides what to do after task execution outcome
• Execution Agent: Executes the instructions using Flow-engineering.

System Level Instructions (Prompts)

• Common/Shared: persona, description_of_processing_layers
• Unique: current_processing_layer, processing_layer_goal, final_output_structure, methodical_steps, processing_layer_rules
• Selection of Abilities: abilities_access_description

Working Memory

https://github.com/markokraemer/mirko.ai/issues/5

Chain-of-Thought / Self-Debate System Logic

A system designed to engage in self-debate recursively up to X steps, ensuring the agent dialogues with itself until it produces the final XML/JSON. Then, it transitions to the next agent.

Layer1

• start_new_thread
• add_message (working memory)
• run_thread
• Loop until we Parse the final JSON —>

Layer2

• start_new_thread
• add_message (working memory)
• run_thread
• Loop until we Parse the final JSON

Layer3

• ..

Communication Between Agents

• Parse final output XML/JSON with next action & deterministically send it to that agent based on the JSON/XML Contents

Technology Choice & LLM Utils setup

• Utilizing the Open Source Assistants API as our agent foundation (for LLM chaining, chat threads, etc.) – this is an open source self-hosteable server implementation of OpenAI GPT Assistants API Refer to the full OpenAI Documentation for more details.
• Decision: Opt for GPT-4 / Claude Opus or both. Using both could mean employing two styles of prompts (XML & JSON) and two parsing methods.

[markokraemer]

Decision: GPT-4 / Claude Opus or both --> both would mean two style of prompts (XML & JSON) & two styles of parsings --> Maybe we go easier route with Claude as everyone has been praising it for code.