The Human in the Loop Still Has to Think
Everybody says: “Keep a human in the loop, mate!” Cool.
What is the human actually doing in that loop?
I have lots of conversations with my LLM about features I wanna build. The majority of my time goes into PRDs, specs, requirements. You name it.
Then I sometimes catch myself doing this: “What would you recommend?”
The LLM gives me a plausible answer. I agree. Next question. Another recommendation. Agree again.
Look at me! Human in the loop. Holy shit! I’m doing it!
Except my thinking has left the building.
Outsourcing thinking to an LLM is a terrible idea. It will always find a solution that sounds good enough.
And we hear that a lot: “Good enough is good enough.”
Until it isn’t.
Constantly agreeing with your LLM leads to:
- concepts you don’t actually understand
- requirements full of slop and vagueness
- software factories producing bad decisions 10x faster
What a bad place to be in.
Dex Horthy—follow him, read his stuff, watch his talks—recently published a long and very good piece about why software factories fail. His argument: more harness engineering, more loops and more review agents can’t compensate for models that don’t preserve software quality over time.
What should I say? I agree!
But I think the factory often fails before the first line of code is written.
It fails inside the PRD. Models are optimized for plausibility, not correctness. If your thinking has holes, the LLM will happily fill them with something that sounds reasonable. Then your coding agents implement it. Perfectly.
Congratulations. You have automated your own confusion.
A conversation with an LLM should create alignment. It should expose the blanks in your thinking. It should force you to make decisions and understand what you’re creating.
An LLM is a mirror of your thinking.
Not the other way round.
A human pressing “YES” is not a verifier. It’s a rubber stamp with a pulse.
You still have to do the thinking. You still have to create backpressure. You still have to own the decisions.
The state machine is more valuable than the agent.
The agent comes afterwards.
Originally posted on LinkedIn.