Two camps, one mistake

I spent last week arguing about AI agents in two opposite directions.

One opponent argued that LLMs can't produce anything genuinely new: they think nothing like humans, they generate text by chance, and therefore they have no place in regulated production. Fit only for narrow tasks with hard rules and brute-force variant search.

The other argued the opposite: agents have already crossed the line where you can hand them a task, tell them to check themselves with other LLMs, and get results more reliable than a human's. His proof: code generation, with unit tests as the referee.

Dismissal in one chat, worship in the other. Two camps that would happily argue with each other.

Except they're making the same mistake.

Both measure the agent on a human scale. One places it below us: a dice-rolling toy, not to be taken seriously, so its risks are ignored as trivial. The other places it above us: a better employee, trust it and step aside, so its risks are ignored as solved. Opposite directions, identical result: the engineering discipline gets switched off.

To the first, I showed that an LLM can be made fully repeatable and auditable, and the risk still doesn't vanish: the answer depends on context, and an unexpected context produces an unexpected answer. Determinism is not safety. To the second, that unit tests are the referee of syntax, not of sense: above them sit user and operational acceptance, and that's exactly where context gaps and answer variability surface. Green tests are not acceptance.

Then something interesting happened. Both stopped defending positions and started designing: what would a solution look like that manages these risks instead of dismissing or denying them? Different camps, same drawing board.

That drawing board has a name. The agent isn't below the human scale or above it. It's off that scale entirely: a different kind of participant, needing its own discipline. That discipline is what my book is about.

Which camp is louder in your organisation: dismissal or worship?