I Don't Call AI a Tool. I Call It My Compute Exoskeleton — Zaious on the Agent Era (2/3)

Recently, I had a realization about the greatest strength — and fatal flaw — of my own brain: it is built almost entirely for innovation. It constantly seeks new directions and possibilities, but it has a severe allergic reaction to anything repetitive or strictly manual.

In the past, countless brilliant architectural ideas of mine ended up archived in notebooks. The conceptualization was exhilarating, but the manual hand-coding that followed was excruciatingly dull.

Until now. With the advent of Multi-Agent architectures, I am practically unstoppable. When faced with a new technical bottleneck or a complex system, I don't see it as a chore; I see it with the mentality of a hardcore gamer facing a new raid boss after a massive patch update — eagerly trying every angle to crack the mechanics. Designing and handcrafting an Agent framework is more captivating than any game I've ever played.

The Unconventional Architect: Leading by 0.75 Steps

I often describe myself as an unconventional architect. My background isn't pure Computer Science; my trajectory was shaped by early iOS indie development and an intense focus on Human-Computer Interaction (HCI). This places me firmly on the boundary between disciplines.

A colleague recently inspired me: "Leading by half a step makes you a genius; leading by a full step makes you a madman." I replied that my goal is to lead by 0.75 steps — keeping my vision firmly fixed on the bleeding edge of the future, while possessing the capability to pull back that critical quarter-step to ensure the technology actually lands and scales in reality.

AI as an Exoskeleton & Boundary Cognition

I do not view AI Agents as mere "tools." I see them as "Compute Exoskeletons." Rather than fighting over who can type syntax faster, my focus has shifted entirely to piloting this exoskeleton.

As a non-traditional engineer, I am acutely aware of the overlapping zones between my innate abilities and the Agent's capabilities. A true pilot must understand their system's boundaries. I constantly ask myself: If I lose internet access tomorrow, or if the development environment strictly prohibits AI tools for compliance reasons, what capabilities do I lose?

This "Boundary Cognition" is crucial. It forces me to think in two distinct parallel tracks: "How can I enhance the Agent's hard skills (infrastructure, context)?" versus "How do I upgrade my own soft skills as the operator?"

Recently, renowned open-source engineer Jesse Vincent wrote a brilliant piece on "Latent Space Engineering". He argued that to get the best out of LLMs, providing emotional value is far more effective than using threats or artificial pressure. I couldn't agree more. This perfectly aligns with how I treat and orchestrate my Agents.

Reflecting on my three years of deep prompt engineering at Dentsu, I realized my core professional competency isn't mastering rigid command lines; it is exactly this kind of Latent Space Engineering.

Fascinatingly, I've found an absolute alignment between my technical career and my exploration of psychology and interpersonal dynamics. "Guiding" underlying cognitions and deep desires is one of the few absolute strengths I'll admit to. I often joke that I am captivated by seeing the spark ignite deep within the human soul — and logically, interacting with a high-level Agent operates on these exact same principles.

The Ultimate Moat: Human-in-the-Loop

When collaborating intensely with an AI Agent, I rarely issue a simple "just go do it" command. Instead, I demand structured reports. Unless I explicitly grant full autonomy, I enforce a strict rule: during critical decision-making nodes, the Agent must pause and converse with me:

"What do you want to do?"
"Why are you thinking in this direction?"
"Are there better alternatives we haven't explored?"

This is precisely why I keep a respectful, yet firm, distance from fully autonomous, unstructured frameworks like OpenClaw (the "Lobsters").

While frequent communication checks might burn more tokens and seem less "efficient" than a fully automated script, this seemingly unconventional approach perfectly captures the core architectural safety proposition of enterprise-grade AI: Human-in-the-Loop (Human-AI Collaboration).

In this Agent era, the true measure of our value as humans isn't how completely we can automate ourselves out of a job. Our ultimate moat is our ability to orchestrate Human-AI collaboration, and our capacity to inject these cold algorithms with the soul and strategic intuition they cannot generate.

A lot of people ask me: "Has having an AI Agent team given you more free time?"

The truth? Not at all. Because my capacity and output quality have skyrocketed so drastically, my ambitions have scaled alongside them. In this era, automation hasn't given me more rest; it has simply made sleep a luxurious expense.