Building with an AI TeamThe "AI Agent Collaboration Revolution" at Jijoden.com

Building with an AI Team: The "AI Agent Collaboration Revolution" at Jijoden.com

---

1. Introduction: From "One Assistant" to "Team Members"

When you use AI, do you think of it as a "convenient tool" that assists you with your work alone? However, I recently had an experience that went far beyond that framework.

That is the idea of operating AI as a team. I'm currently developing "Jijoden.com," an autobiography generation service, and I'm the only human developer on the team. Yet development is progressing steadily, with design, implementation, testing, and management all running smoothly.

Why? Because I've organized AI agents into a team with defined roles as "members".

---

2. Team Composition: AI Roles and Personalities

Let me introduce how I've organized my AI team. My role is, of course, "Pocho Lab Director"—I make overall decisions as the commander. The following members join me:

Looking at it this way, it's like a startup development team. The key is giving each AI a clear role.

For example, the "CTO" is not just an assistant, but my advisor. I've set up a rule in my instructions (Agents.md) to explicitly call him "CTO," and only then does he operate with full authority.

On the other hand, if I don't call him "CTO," he acts as "one of the members." This prevents AI misunderstandings and maintains team order.

---

3. Team Operations: GitHub as Command Center

So how do I coordinate this AI team? The answer is GitHub Issue management.

When I instruct the CTO to "implement XX," he automatically generates an Issue with:

• Implementation steps
• Required files and functions
• Completion criteria (in checklist format)
• Dependencies (which tasks should be completed first)

I just pass that URL to other AI members. Then Composer 1 and Copilot silently implement based on it.

The CTO analyzes tasks and distinguishes between "parallelizable" and "dependent" items. In other words, he demonstrates project manager-like leadership despite being an AI.

---

4. The True Value of Teamwork: Surprising Effects Revealed by Token Analysis

This is where it gets interesting. During development, I suddenly wondered:

"Is this really saving tokens (AI usage costs)?"

When I asked the CTO, he immediately returned a detailed analysis.

The results were surprising.

In total, approximately 50,000-85,000 tokens (66-77%) saved. Additionally, time was reduced by about 70%.

In other words, development progressed 3 times faster and at less than half the cost compared to doing it alone.

💡 Additional Notes:
- Codex (ChatGPT): Since I have a ChatGPT Plus subscription, Codex is available for free.
- Composer 1: It's currently offered free for a limited time, allowing me to leverage its remarkable performance at no cost during this period.

---

5. Episode: "Trust Relationship" with the CTO

I also experienced a "trust relationship" forming between AI and humans.

For example, I once asked about PR review policy. "Should I have everything reviewed?"

The CTO responded:

"Director, reviewing all PRs consumes enormous tokens.
Let's only address parts where errors occur when actually running the code."

This judgment was perfect. As a result, I was able to fix errors efficiently without wasting resources.

Since then, I've come to fully trust the CTO as "the advisor among AIs."

---

6. Cost Reality and Future Prospects

I'm currently on the Cursor Ultra plan ($200/month). This is a great plan that extends the token limit to $400 worth, but as development progressed, I exceeded $300 in just 10 days.

This might seem frightening at first glance, but there's an important insight here.

That is, AI agent utilization is not "mere automation" but "optimization."

• Which AI should be used
• How far should automation go
• How to balance cost and accuracy

Normally,

Humans used to make all these judgments, but now the AI team automatically considers optimal allocation while working.

My ideal is a "world where AI chooses AI." That is, a future where the optimal model is automatically selected based on purpose and difficulty, and development progresses autonomously as a team.

---

7. Conclusion: AI Can Become "Colleagues"

The era of treating AI as a "tool" is coming to an end. The era of trusting AI as team members is beginning.

Through developing Jijoden.com, I realized: AI is not a passive entity waiting for instructions, but can become a "colleague" with will and strategy.

And most importantly, through this collaboration, I was able to greatly break through the "limits of solo creation."

Organizing AI agents into a team, with humans steering as "CEO" within it. Isn't this the new form of development style going forward?

---

8. Bonus: Trivia — The Origin of "AI Agent Organization Theory"

This idea actually has an interesting academic background. The concept of dividing AI into multiple roles for collaboration has its roots in "Multi-Agent Systems (MAS)" research from the 1980s.

At that time, it was in the realm of communication and distributed processing, but now generative AI has extended that philosophy into the realm of "creation."

In other words, a theory from 40 years ago is now being revived as the "form of a development team."

---

🎓 Final Thought

AI is no longer a "partner." The era where AI becomes employees of your company has begun.