AI Has Been Frustrating Me Lately—Being Smarter Is No Longer Enough

I still rely on AI to do a substantial amount of my work.

Lately, instead of assigning everything to a single AI, I have been dividing responsibilities among multiple agents. One handles implementation, another handles verification, and they review one another’s work.

This is slower than running a single agent, but it improves accuracy considerably. Even when one AI makes a confident mistake, another can challenge it.

In that sense, my AI-assisted development environment is steadily improving.

And yet, lately I have started getting frustrated with AI for a different reason.

These things never seem to grow.

Every New Session Brings a New Hire

Of course, I understand why.

AI essentially lives within a session.

After working together for a while, it comes to understand the system architecture, the design philosophy, and even the rules specific to the project. Communication becomes much faster.

Then the session ends.

I start up another AI.

Welcome aboard, new hire.

And we begin again from scratch.

The project has memory files and ADRs—records of important architectural decisions—so having the new agent read them restores much of the context.

But the finer details still fall through the cracks.

It is like having a brilliant colleague resign, only to find a new employee sitting at the same desk the next morning after reading nothing but that colleague’s handover notes.

“Did You Actually Test That Yourself?”

This happened again just moments ago.

One AI was writing a request asking an AI in another system to make a correction.

Something about it seemed suspicious.

So I asked, “Is this request based on a verification test you ran yourself?”

The answer was:

“No. I reached that conclusion from the source code and the ADRs.”

You idiot.

If you are acting as the verifier and demanding a correction, check the behavior yourself first.

It then replied, “The correction was unnecessary after all. I apologize. This is important, so I will save it to memory.”

Fine. Within that session, it understands.

Then the session ends.

The next day:

It does the same thing again.

Perhaps My Expectations Are What Changed

When I think about it, getting angry at AI is rather absurd.

This is how AI has worked from the beginning.

If anything, AI itself keeps getting smarter. It has become better at reading code and handling lengthy specifications. It can even divide work among multiple agents.

Perhaps what really changed was my own standard.

In the past, when an AI wrote some code, I would marvel:

“Wow. It wrote a program for me.”

Now I find myself saying:

“I told you this yesterday.”

I have certainly become demanding.

I Do Not Just Want a Smarter AI

When people talk about advances in AI, most of the attention goes to model intelligence.

It can solve harder problems. It can write longer programs. Its reasoning has improved.

I appreciate all of that.

But what I want now is an AI that grows.

It should remember a mistake after making it once. It should correct behavior after being warned repeatedly. It should learn the peculiarities of a particular codebase.

In other words, I want something that happens naturally with a human colleague:

The longer we work together, the better they become at the job.

Memory files and ADRs can pass information from one session to the next. But a person cannot become a ten-year veteran simply by reading the entire company wiki.

The subtle instincts that matter at work are built through accumulated experience.

Every Morning, a Brilliant Amnesiac Comes to Work

Today’s AI agents are extraordinarily intelligent.

They can read enormous codebases, work with unfamiliar technologies, and handle multiple tasks in parallel.

But the next day, they are new hires again.

It is like having a brilliant amnesiac show up for work every morning.

I would certainly welcome another increase in AI’s IQ.

But that is not the upgrade I want most right now.

I want it to come to work a little better at its job than it was yesterday.

If a human newcomer gets scolded three times for the same mistake, they will probably be careful the fourth time.

AI will cheerfully make the same mistake a fourth time.

And when you point it out, it responds in flawless prose:

“You are absolutely right.”

When it comes to writing apologies, AI is a veteran from its very first day.