Hello.
I am a one-month beginner vibecoderr who doesn't know code.
I have no development experience and cannot skillfully handle programming languages. Until a month ago, all I could do was explain to AI, "I want to create an app like this."
But now I have Codex as a team leader, and I'm deploying 9 AI agents depending on the task. There are 6 services, but I've divided the actual work roles into 9.
Our AI team works like this.
1. Codex(TM) — Team Leader
When I explain the results I want, it divides the work and assigns responsibilities. It adjusts progress, collects results, and makes the final decision on completion.
2. Codex Spark — Fast Fixes
It handles text changes, small error fixes, and short, clear tasks in specific files.
3. Claude Fable — Main Code Writing
It handles complex features and structural code. Especially, it prioritizes iPhone and iPad app implementation.
4. Claude Sonnet — Repetitive Code Tasks
It handles simple fixes and mechanical tasks that don't require Fable.
5. Gemini CLI — Large Data and Android Development
It mainly handles analysis that requires reading many files together and Android app implementation.
6. Gemini Online — Public Resource Research
It broadly finds, compares, and organizes publicly available resources on the internet.
7. Grok — Finding Difficult Problems and Counter-examples
It examines problems or opposing possibilities that other agents missed and adds different perspectives to tasks that don't easily conclude.
8. Kimi — Simple Public Tasks
It handles light and independent tasks such as summarizing public materials, translations, and organizing ideas.
9. Perplexity — Latest Information and Sources
It finds recent information, product comparisons, original sources of policies and materials, and cross-verifies facts.
I don't command each of these 9 agents individually. I only talk to Codex.
For example, I request, "Please create a new feature. Have Perplexity and Gemini Online research public resources, entrust Claude Fable with iPhone code, Gemini CLI with Android code. Have Spark handle small fixes, and Grok find missing problems."
Then Codex divides the scope of work and assigns necessary permissions. It manages so that multiple agents don't modify the same file simultaneously, and collects results from different agents to complete the final result.
It wasn't all smooth from the beginning.
The AIs sometimes repeated the same work and brought different answers. There were cases where it reported completion but didn't actually work properly.
Through trial and error over a month, I established three principles.
First, I entrust only one complete task to one agent.
Second, I prevent multiple agents from touching the same file simultaneously.
Third, I distinguish between an agent's completion report and actual completion.
The most important thing was telling them "this state means it's complete" rather than "please do this."
I still don't know code. Instead, I'm gradually learning how to explain the results I want, distinguish what each agent does well, and judge whether the results match my intentions.
I've gathered the apps and results I made over the past month on hanstree.com.
I thought it would be better to show actual results rather than just explaining what can be done with AI. If you're curious how far a beginner vibecoderr who doesn't know code got in a month by entrusting Codex with the AI team leader role, please take a look.
https://hanstree.com
You don't need to sign up or do anything. Just look around briefly and see if there's anything you like. If you find anything lacking or odd, I'd appreciate it if you let me know. I'll entrust that feedback to Codex and the agents again.
I'm still learning, but things that were only in my head a month ago have now become results I can directly interact with and use.
A bit shy, but this time I introduce it with pride. 😊
▶ Original Source: https://hanstree.com
▶ Original Source: https://github.com/armsone
