We’ve once again taken on reworking the personal assistant based on OpenAI Codex — or, to put it simply, like shooting sparrows with an AI cannon, trying to burn out maximum efficiency. This is already the second part of the experiment.
So, like many, I continue experimenting with the idea of a personal helper that can go online, run code, and manage files in folders. The engine I’ve chosen for these experiments is OpenAI Codex — mainly used with a hyper-expanded GPT-5.2 in High Reasoning mode, without much hassle.
I spent a week relying on the structure I created at the start (if needed, see the context). In my GitHub repository, I have this scheme:
├── 00_inbox
├── 01_capture
├── 02_distill
├── 04_projects
├── 07_rfcs
├── 89_images
├── 90_memory
├── 99_process
├── AGENTS.md
└── CLAUDE.MD
Inside — markdown files on projects, task cards, transcripts of conversations, and other notes. This repo can be opened in Obsidian, Codex (locally or via cloud), or Claude Code. Maintaining it manually is tedious and long-term; automation does everything for you here.
For now, I used this repository as a central collection point for all information passing through me — it was very convenient to send a stream of thoughts via voice after meetings or conversations: like, “here’s the link and ideas.” Codex (also available on phone) quickly picks up the command “make capture,” and then everything proceeds according to a preset scenario — cards are created automatically to make it easy to find what’s needed later.
In fact, I usually just say “make capture,” and then the instructions handle everything themselves. All this works because OpenAI Codex is a good alternative to more complex systems like OpenClaw or MoltBot. No setup needed: all components are already available and work out of the box.
The information capture process has become very effective. However, autonomous agents have not yet shown the same results. Over the week, I developed a decent solution: a quite useful personal librarian.
But the main idea of this experiment is not to stop there. I want to find new ways to use agents more efficiently and develop structures for automatic data capture. So today, I decided to go further and integrated SOUL_MD from OpenClaw into my repository — with the ability to change any files in the `90_memory` folder without my permission. After that, I ordered a scenario version for automatic task execution.
Doing this safely means there’s nothing to lose: if something goes wrong, I simply roll back git to the current state.
And then the most interesting part began. Codex not only devised a concept for an agent management center and a task tracker (issue tracker), but also amusingly tinkered with executing its own tasks. It extracts insights from calls or meetings and turns them into ideas for improving its own processes.
So far, the benefits are mostly theoretical — nothing groundbreaking has been invented yet — but I already have a couple of ideas on how to build agent workflows (so-called “context graphs,” as they say nowadays in the SV and OpenAI worlds).
I plan to attach some screenshots of this system in comments — showing SOUL_MD and how it works.
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby