DEV Community

lingxin wang
lingxin wang

Posted on

The Modular Mind

#ai #agents #architecture #llm

Why AI Capabilities Can Be "Bought"

A founder friend asked me last week: "You're working on that AI Agent project, right? I want to build an assistant that automatically queries data for customers. Do I need to spend a lot of money hiring someone to write the code?"

I said: "Not necessarily. You just need a few functions."

He didn't believe me: "An AI that can query databases? Just a few functions?"

I explained it to him. When I finished, he said: "The logic is kind of like a hospital."

He nailed it.

How Hospitals Work

Think about how hospitals are organized.

A general clinic — the doctor has to handle colds, fractures, and simple surgeries. One person does everything. Long training period, high replacement cost.

But big hospitals don't work that way.

Big hospitals have triage, internal medicine, surgery, labs, radiology... each department specializes in one thing:

  • Internal medicine handles diagnosis
  • Labs handle testing
  • Radiology handles imaging
  • Surgery handles operations

Doctors collaborate through "consultations": internal medicine says "need a CT scan," radiology goes and does it. Radiology says "results are ready," internal medicine reviews them.

Each department is a module. Modules collaborate through standard interfaces.

AI Agents work the same way.

Agent Capabilities Are "Assembled"

The Agent I implemented has this structure:

Agent(
    model=self.model,              # brain
    tools=[                        # skills
        self.get_database_schema,  # view database structure
        self.execute_sql_query,    # execute queries
    ],
)
Enter fullscreen mode Exit fullscreen mode

model is the brain — understands user questions, decides what to do.

tools are skills — each skill does one specific thing.

Completely separate.

Just like a hospital:

  • Brain = attending physician, responsible for judgment and decisions
  • Tools = departments, responsible for specific execution

Attending says "need a blood panel," lab does it. Says "need an ultrasound," ultrasound department does it. The doctor doesn't need to draw blood or operate ultrasound machines — they just need to know when to call which department.

Agent is the same. It doesn't need to "know how to" execute SQL — it just needs to know when to call the execute_sql_query tool.

Why This Matters for Founders

My friend got excited here: "So I don't need to build an AI from scratch?"

Right. You just write a few Tool functions.

Want Agent to query customer data? Write a query_customer function.
Want Agent to send emails? Write a send_email function.
Want Agent to place orders? Write a create_order function.

Each function is a "department." Agent's brain decides when to call which department.

You're buying "skill modules," not buying an entire AI.

Same logic as hospitals. You don't need to train your own radiologists — you can outsource to an imaging center. You don't need your own lab team — you can send samples to third-party labs.

AI is the same. Brains (GPT, Claude) are already trained by others, you use them directly. You only develop the "departments" (Tool functions) specific to your business.

The Real AI Revolution

Then my friend asked a deeper question: "So what's AI's biggest change? Getting smarter?"

I thought about it and said: Not getting smarter — skills becoming tradeable.

Human skills are "grown" inside brains.

A doctor spends 10 years learning diagnosis. They can't "copy" that skill to someone else.
A chef spends 20 years perfecting knife work. They can't USB-transfer it to an apprentice.

Skills are bound to people. Can't circulate.

But AI skills are "plugged into" the brain.

I wrote an execute_sql_query function. Every Agent in the world can use it.
Someone wrote a send_email function. I can just take and use it.
Someone developed a suite of medical consultation Tools. I can "buy" them and plug into my Agent.

Skills become assets that can be copied, purchased, and combined.

That's the real paradigm shift.

Advice for Founders

If you're thinking about building AI products:

Don't think about "building an AI" — think about "buying brains, building departments."

  1. Brain (LLM) — use existing ones directly. GPT, Claude, Gemini all work
  2. Your moat is your Tool functions — these are your "specialized departments"
  3. Each Tool does one thing well, can be tested and iterated independently

This architecture has low cost, fast iteration, small risk.

You're not building an AI from scratch. You're assembling a "specialty hospital."

Personal Reflection

As someone fresh out of school, I've learned more than just technology from this design.

I used to think impressive people "know everything."

Now I think impressive systems are "modular" — each part does one thing, combined they're incredibly powerful.

People are the same way. Instead of chasing "know everything," maybe cultivate yourself as a "pluggable module" — excel at one thing, then combine with other excellent modules.

That's probably a career insight learned from AI architecture.

Top comments (0)