Several engineers on our team describe working with a coding agent as "like pairing", and then get frustrated when it does not behave like their usual pair. It is worth being precise about what is actually different.
A human pair pushes back
A colleague who thinks your approach is wrong says so, usually before writing any code. A model, by default, does not push back — it produces its best attempt at what was asked, even when the request rests on a bad assumption. Treating the two as equivalent means losing the single most useful thing about pairing: someone catching a mistake before it is written.
What we do instead
We ask the model to state its plan before implementing, and we read the plan the way we would read a colleague's proposal — for the assumption buried in it, not just the steps. This recovers some of the pushback, because a stated plan is easier to disagree with than a finished diff.
Turn-taking is different
Human pairing alternates who is typing, and the person not typing is thinking one step ahead. With a model, the equivalent discipline is deciding, before the model starts, what you will check for when it finishes — not watching the output scroll past and hoping something looks wrong if it is wrong.
- Before: state the acceptance criteria out loud, even to yourself.
- During: do something else. Watching an agent generate code is neither review nor rest.
- After: check the stated criteria against the diff, deliberately, before reading the code for style.
The skill that atrophies if you are not careful
Debugging someone else's unfamiliar code is a skill, and it is the skill this way of working exercises constantly, because agent-written code is always someone else's code, even on the day it was written. Engineers who skip straight to "does it pass" without reading the diff are quietly losing the skill that catches the next serious problem.
The point of pairing was never that two people type faster. It was that two independent judgements catch more than one. A model gives you a second pair of hands. It does not give you a second judgement unless you insist on it.
Where it genuinely helps most
Boilerplate, migrations, and the tedious middle of a refactor that a human pair would find just as tedious. The agent does not get bored, which is a real advantage on the kind of work where a bored human starts making the mistakes that boredom causes.