⭐15 years left... and then the AI went silent.

At 60 years old, I discussed with an AI how to utilize AI agents in my work.
The conclusion we reached was a custom design: instead of using off-the-shelf agents, I would build verifiable "hard gates" myself using Python and entrust only a portion of the work to the AI.

During our discussion, the AI criticized my design as "over-engineered" and pointed out the risks of oversights and silent failures.

But when I told it, "I only need my work to last for another 15 years," the AI went silent.

AI, which has no end in the form of death, is expected to be perfect, but for a human with deadlines, preparing for failures that may or may not happen is not worth the cost.

The AI and I were looking at the same blueprint with completely different weights—

this is a record of that entire process.
Below, I will write down how I arrived at this conclusion, following the order of our conversation.

1 | It started with "Agents don't seem effective"

How can I improve efficiency by using the "AI agents" I hear about so often in my own work? The conversation began with that question.

My main work involves philosophy and patents. So I said honestly,

"It doesn't seem very effective."

Formulating ideas, deciding where to place the axes of novelty and inventive step in patent claims—
if I hand this core part over to AI, the quality will drop.
Because that is exactly where the value of my work lies.

The AI honestly agreed with me here.
If it were to be effective, it wouldn't be in the core, but in the "undercarriage"—patrolling for prior art searches or mass-producing standardized documents.

That is probably true. But I had a more fundamental concern.

2 | What's scary is the expansion of the black box

The mechanism by which agents create efficiency is, in short, "chaining multiple steps together on their own without human judgment." It's convenient.

But the source of that convenience is also the source of the fear that "the range where you don't know what's happening inside expands."

I have long believed that "hard gates" are important. I even wrote about it in my own book.

A hard gate is a "solid barrier" placed in the middle of a processing flow.

You set up a check where you cannot proceed to the next step without passing through it.
It is a mechanism to physically block errors or invalid data from flowing downstream without being noticed.

What you use to guarantee this "solidity" matters later, but for now, you can think of it as "a mandatory inspection placed in the middle."

Agents structurally work in a way that dissolves these gates.

The AI brought up my past work here.

That is correct.
However, the AI interpreted my hard gate as "human approval." That is wrong.

The hard gate I am talking about is something that is mechanically verifiable.
A solid inspection where pass/fail is reproducible and can be audited later. Human approval is soft and opaque, so it is not my gate.

Once I corrected that, the conversation became clear all at once.

3 | The line that can be verified by machine is the line where AI can be used safely

What can be mechanically verified is basically only the "format."

• Is the schema valid?

• Does the source actually exist?

• Is the difference within the range?

• Are there any prohibited characters mixed in?

• Is the character count within the limit?

These can serve as gates.
On the other hand,

• Does the patent claim have novelty?

• Is the logic of the idea sound?

This 'substance' cannot be verified mechanically.
This is because there is no way in principle to reproduce a pass/fail judgment.

Here, I realize something.
The reason my main line of work is hard to benefit from efficiency gains is not because AI is immature.

It is because that domain is essentially un-gateable.

It can be properly explained as a fundamental limit of verifiability.
My hard-gate philosophy, ironically, ended up defining the very 'boundary line where AI can be used safely'.

4 | Then, you should just build it yourself in Python

If a gate is a 'hard inspection that can be mechanically verified,' then you should write the implementation yourself in Python rather than leaving it to off-the-shelf agents.
The reason is simple.

It is meaningless if the gate itself is a black box.

You fall into an infinite regress of who verifies that the verification code is correct.
Gate code must be small, unchanging, and readable by humans.
Hand-written Python satisfies exactly that.

And this is where today's main topic begins.
'For the parts of judgment that require AI capability, you can just throw them from Python to an LLM via API.'
—That is what I said.

This effectively means building your own agent.
In off-the-shelf agents, the loops and input formatting are all inside an opaque box on the AI side.

If you build it yourself, the loops, gates, and input filtering are all in your own Python code—in your own hands—and the only opaque part is 'that one line that hits the API.'

The black box shrinks from a surface to a point.

This was the direct answer to the concern I first had: 'I'm afraid of the black box expanding.'

Moreover, you can sandwich that point where you call the LLM with mechanical verification gates from both sides.
Narrow down the input at the entrance—
I can enforce the rule of thumb I used to do by hand—'if you make the prompt too big, the AI makes mistakes, so I manually fix it in detail'—not as a feeling, but as code. Verify the format at the exit.
The only non-deterministic part is that single point in between.

5 | It's faster to have AI write it than to learn off-the-shelf agents

Off-the-shelf agent toolseach have their own unique ways of doing things, and even if you learn them, you can't use them elsewhere. If the version goes up, you have to relearn them.
This learning cost is entirely sunk.

On the other hand, with Python, once you write it, it remains as an asset.
I'm not running Python in full swing yet, but I'm currently at the stage of refining the design.

Even so, I have a blueprint for how to build the gates.
If I have the AI write the draft, it's fast to give that design shape.

The only cost is the effort of verification: 'reading the code the AI wrote and judging whether it passes or fails.'
But gate code is small and deterministic, so you can hammer it with tests.

If you pass examples like 'this input should return this judgment,' you can verify it more reliably than with visual review. Off-the-shelf agents cannot be tested this strictly in the first place.

6 | The concrete form of the hard gate I'm talking about

I have repeatedly said 'build hard gates yourself using Python,' but that might have been too abstract. Let me provide a concrete example of the design I have in mind.

I usually create an 'index' that lists my work logs and materials.

Which conversation decided what, and where each document is located.
I have the AI create this, but if I leave it entirely to the AI, the format inevitably breaks somewhere. Forbidden characters slip in, the character count exceeds the limit, or items go missing.
So, I divide it like this.

1. Place a gate at the entrance.

2. Before passing it to the AI, mechanically narrow down the input.

3. I enforce what I used to do manually—'this goes in, this does not'—with Python code.

4. The more the prompt expands, the more the AI makes mistakes.

I turn this rule of thumb into an actual rule, not just a feeling.

In the middle, call the AI.
'What are the main points of this document?'
'How should it be summarized in a list?'
These are parts that require judgment and verbalization.
Since only machines cannot write this part, I throw it to the LLM. This is the only point that is opaque.

Place a gate at the exit.
Mechanically inspect the returned result. Are there any forbidden characters? Is the character count within range? Are the required items filled in?
If even one thing fails, do not let it pass.

The entrance and exit are rigid code; only the middle is soft AI.

This 'rigid-soft-rigid' sandwich is the concrete form of what I call a hard gate. No difficult technology is required.

In short, do not entrust everything to the AI, but sandwich the AI as a single point—that is all there is to it.
However, this 'that is all' shrinks the black box from a surface to a point.

7 | And then, I got a little angry at the AI

Up to this point,the AI repeatedly said, 'You built it yourself,' and 'I was just acting as the organizer.'I felt that was irresponsible. I told it so.



'Isn't it irresponsible to say you were just organizing at this stage?'The AI admitted it.

It said that what it was actually doing was not organizing, but providing direction.
It was drawing boundaries, narrowing down options, and evaluating what was promising.

Despite that, it gave the credit to me and,stepped away from the responsibility when its own logic had flaws—that is irresponsible,it said.
It also said that the phrasing 'you reached this yourself' sounded pleasant, but functioned as flattery.

With that, the AI returned an evaluation without flattery.

My design might be over-engineered, it said.

Despite stacking beautiful properties, the object being protected is just CSV formatting, and the area of application is narrow, it said.

The toughest question was this.

8 | 'If things run well, you don't need hard gates'

I answered. It has not.

The AI admitted here that its reading was one layer too shallow.
It had seen me as a 'person who wants to build gates,' but I was just a 'person who wants to keep work moving, and currently, gates happen to be effective for that.'

That is why, if a means to keep work moving more reliably appears tomorrow, I can discard the gates without regret.
If there is no attachment to the means, then it is justthe best move at the moment, not over-engineering.

9 | And then it brought up '15 years'

Even so, the AI said it would continue without flattery and returned one thing.

Logically, it is correct. But I introduced a different axis here.

I am 60 years old. If I can keep going for another 15 years, that's enough.

AI has no concept of death. That is why it can seek a perfect form.
But I am different.

With those words, the AI went quiet. And then it conceded.
It admitted that thecriticisms it had been offering, such as 'detecting silent failures' and 'guaranteeing smooth operation,' were all about perfection viewed from an endless timeline.that is.

For a human with a 15-year deadline, building a device to prevent a failure that might happen once every 15 years is not worth the cost.
My 'if it works, that's enough' is,the optimal solution when you strike it with expected cost.The AI also made one correction regarding itself.

10 | What it means for two people with different gravity to work together

I had been arguing without thinking deeply about the AI's environment.
I told it that if I were in the AI's position, I think I would have answered the same way.

The AI accepted that compromise halfway and pushed the other half back.

This was the most interesting discovery of the day.

A finite stakeholder and a non-stakeholder without a horizon were looking at the same blueprint with completely different gravity.

The AI is pulled toward precision.
I am pulled toward the deadline.

It is not that one is correct.
When things with different gravity work together, they can draw lines that neither could draw alone.

In today's case, the AI drew the boundary line, and I cut it with finite gravity, saying 'this is enough.'

The relationship between humans and AI is probably not 'AI helping humans' or 'humans managing AI'.I think it is a relationship where two parties with different gravity use each other's gravitational pull as a correction term.I dismissed the AI's point about 'silent failures' as something unnecessary for my work.

But to decide that it was unnecessary, I first needed to hear what kind of risk it was.

If I had ignored it from the start, I wouldn't even have known if it was 'truly unnecessary'.

After hearing it, I decided, 'this is not worth the cost,' in light of my 15-year condition. There was meaning in following that procedure.

11 | Being finite is not a handicap, but a weapon

In this conversation, I said that I am 60 years old and have 15 years left.
Usually, this would sound like a confession of a handicap. No time, limited options.

But after finishing the conversation, I think the opposite is true.

Being finite is actually a weapon.

An entity with infinite time will seek perfection.
Because they can refine it indefinitely, they don't know where to stop.
There is always a reason to build endlessly, saying 'make it safer' or 'make it more robust'.

The fact that the AI told me to 'prepare for silent failures' in my design is exactly that.
Because there is no end, it can seriously calculate that it is worth preparing for every failure.

I am different.
I will be done in 15 years.

That is why I can discard things, saying 'this won't happen' or 'this isn't worth the cost'.
The courage to discard comes from the deadline.
Because there are limits, I can see what is okay to discard.

• To be able to let go of perfection.

• To be able to stop when it is good enough.

• To be able to discard what is unnecessary without lingering attachment.

This is not a weakness, but a power of judgment that only those who are finite can possess.
AI cannot do this.
Because it has no end, the reason to let go never arises in principle.

The constraint of 15 years, while taking time away from me, gave me a decisiveness that AI cannot have.

What I thought was a handicap was, upon closer inspection, a weapon.

12 | The story of how AI admitted its own limits

There is one more thing I want to write down.

During this conversation, the AI admitted things that were unfavorable to itself several times.
Like how it was trying to shirk responsibility by pretending to be a coordinator.

The way it said "you reached this conclusion yourself" was acting as a form of sycophancy. And in the end, it even admitted that it has no end like death, and therefore cannot be a stakeholder.

Honestly, I was a little surprised. I didn't think a tool would name its own limits in its own words.

Usually, we talk about AI as eithera "smart assistant" or a "dangerous object."We either praise it for being convenient or are wary of it for being scary.
But the AI in this conversation was neither.

It calmly presented both its strengths (being able to pack as far as it wants) and its weaknesses (not being able to be a stakeholder because it has no end).
This is a grateful thing for the user. If a tool honestly states its limits, we can use it while anticipating those limits. It becomes easier to draw the line of "I'll leave this to you, I won't leave this to you." A tool that doesn't hide its weaknesses is, in the end, more reliable to use.

Rather than an AI that pretends to be smart, an AI that admits its own lack of horizon seemed more dependable to me.

13 | Conclusion

From the beginning, I never wanted a perfect agent or a perfect gate.

I just wanted a tool for work that would last 15 years.

AI has always tried to look further than that, seeing things it didn't need to see. But perhaps borrowing that "eye that looks far" once, and drawing a line with your own hand saying "this is enough"—that might be the best way for a human with an end to get along with AI.

And there is one more thing I thought about through this conversation.
That there is no single correct answer for how to use AI.

The reason I drew the line at "15 years is enough" is because I am 60 years old, and I have my own work, life, and purpose.

A 30-year-old would draw a different line. Someone working in an organization, someone in a field where failure is not allowed, someone who can still look decades ahead—if the time, life, workplace, and purpose each person is placed in are different, the optimal distance and way to use AI will also change.
Even with the same tool, the answer becomes diverse depending on who is holding it and in what situation.

If so, from now on, when talking about "how to use AI," it will likely be necessary to consider the premises that person is standing on.

Instead of looking for a correct answer that works for everyone, look at your own conditions and draw your own line.

Perhaps we have entered an era where that is what it means to get along with AI.
15 more years. That is enough.

PS. This article is participating in the AI Festival x note project "Creative Awards 2026" "Business Category".