I’m not sure if I’m in the right category.
Hello, I’m writing this to share some deep thoughts with fellow AI enthusiasts.(I’m not an expert in AI.)
When we do complex work, humans naturally break it down into steps: ‘data collection → analysis → outlining → drafting’. I decided to try applying this simple principle to AI. This was the start of an experiment with an AI architecture written in human natural language, which I call the Code of Conduct.
What is the ‘Code of Conduct’?
In a nutshell, it’s a blueprint written in natural language that defines the thought process and principles of action an AI should follow when performing a complex task.
It breaks down a large mission into smaller pieces, just like humans do: Phase → Stage → Task.
It assigns clear roles to the AI, such as a ‘Planner’ and an ‘Executor’.
The entire process of thought and execution is transparently recorded in the file system (workspace), forming a ‘Chain of Execution’.
The key point is that all of this is designed and controlled not with code like Python, but with the natural language we use.
Insight 1: Natural Language as a Backend System Language
The biggest takeaway from this experiment was glimpsing the potential of a ‘natural language-based backend system’. I designed the entire operational process like a conventional system, but instead of using a computer language like Python, I used the human language I know best. Then, I had the AI follow this process to run the system. To my surprise, I confirmed that the entire system operated in the same way as one built on traditional, code-based computer languages.
Insight 2: A Step Closer to AGI
I didn’t initially set out to mimic an artificial neural network. However, once I completed this workflow, I realized its structure took on a shape that was strikingly similar to an artificial neural network. The process where an external instruction (input) is transformed through multiple stages of planning and execution (hidden layers) to produce a final result (output) was chillingly familiar.
The ‘Code of Conduct’ forces a thought process on the AI, telling it “think like this” instead of just “do this.” This structure guides the AI away from the temptation of providing immediate answers and toward deeper reasoning. As a result, I was able to achieve a level of depth greater than or equal to a junior expert in general-purpose domains. As the guidelines containing my know-how were continuously learned (updated), the quality of the output improved noticeably. I dare say, it seems possible to achieve results surpassing those of the latest services like Skywork, Perplexity, and Genspark.
Insight 3 (The Ultimate Idea): A Recursive Agent Society
What if we take this one step further? What if we replace the ‘Executor’ that currently runs a Task with another child agent (A2A, Agent-to-Agent) that follows the exact same ‘Code of Conduct’? This mirrors a human organization where a manager (parent agent) delegates work to a specialist (child agent). If we design this network structure to deepen in layers, like a 2-tier or 3-tier fractal, I imagine it could enable far deeper and more complex reasoning. Furthermore, if the output of one process could be chained to become the input of the next, like connecting train cars, it seems the automation of much more complex tasks would become possible.
However, the Practical Problems are Also Clear.
Despite this rosy future, I’ve hit two major walls in the current Gemini CLI environment.
Excessive Token Consumption: Currently, Gemini CLI doesn’t support A2A, so a single AI model has to play all the roles—Planner, Executor, etc.—by itself. Because of this, all outputs from the previous step become the input context for the next, causing the input tokens to snowball. I sincerely hope that A2A, where each agent operates independently, is supported soon.(I hope that issue #5000will be resolved soon.)
Insufficient Output Token Limit: Strangely, the output token limit in Gemini CLI seems much tighter than in the standard Gemini chat interface. It’s woefully insufficient for capturing a complete ‘thought process’. I was forced to abandon the granular Stage level and expand the scope to the broader Phase level to enforce the process. Even so, it still feels lacking.
Thank you for reading this long post.
The ‘Code of Conduct’ is still a rough and imperfect experiment, but it has been an invaluable experience that allowed me to see AI from the perspective of an ‘architect’ rather than just a ‘user’.
If you’ve tried something similar or have any thoughts on this idea, please feel free to share. Your feedback would be a great help.