The defaults that looked decided
I was doing close-out QA on a feature that was already live when I noticed an edge case whose implementation was wrong but looked like a deliberate choice. So either that was a real decision, or something else had filled the gap.
How it got there
I asked the agent to trace it. I wanted to know if this was a locked product requirement, review feedback, an agent default, or old behavior.
The feature contract was vague about that edge case: it didn’t enumerate what each output should be.
The agent filled in the gaps (of course that’s what LLMs do 😊). Every unspecified blank got a reasonable-looking default, because there was no concrete decision to audit against.
What I changed
I had to rethink the edge case, then reopened it with product. That conversation turned up consequences of the original work I hadn’t considered. It led to changing the data shape of the original implementation: a clearer edge case, rather than painting over a default.
The lesson
The dangerous agent output is not the obviously-wrong change you catch in review. It’s the unspecified blank filled with a plausible default that reads as decided.