How do you choose which work to pair on rather than pairing uniformly?

## Why "pair on everything" and "pair on nothing" are both wrong Uniform pairing is a simple policy and that is its whole appeal. It also spends two people's attention on changes that one person could do correctly while half-asleep, and it burns out teams, because sustained pairing is far more intense than solo work. The opposite default — nobody ever pairs — produces its own bill later: areas that exactly one person can safely change, newcomers who take months to become useful, and decisions that get discovered rather than made. So the practical skill is selection, and selection needs criteria you can state before the work starts. ## The screen: risk times unfamiliarity **Risk** here means the cost of getting it wrong: how hard is it to reverse, how wide is the blast radius, how quickly would a mistake be detected, and how bad is a mistake for the people on the other end of the system. **Unfamiliarity** means how much of what is needed is not already in one head: an unmapped area, an unstated requirement, a design with several defensible shapes, a technique new to whoever is doing it. High on both is where a second person is cheapest relative to what they prevent. Low on both is where the second person is a passenger. ### Pair or ensemble by default - **Hard to reverse.** Data reshaping, an interface other teams will depend on, anything where the recovery path is long. - **Wide blast radius.** A change on a path everything else flows through. - **Bus factor of one.** The area has a single person who understands it; pairing is the cheapest way to stop that being permanently true, and the point is to put the *other* person on the keyboard. - **Design genuinely open.** Several shapes are defensible and the choice is expensive to change later. - **Stuck investigation.** When someone has been chasing an intermittent timeout on a hospital appointment scheduler's slot-reservation path for three days, adding hours has stopped working; what is missing is a second model of the system. On an 11-person team that is also the moment to consider widening from a pair to a small ensemble, because the knowledge is spread across several people. - **First change in an area.** A newcomer's first change anywhere unfamiliar, driven by them. ### Solo by default - Well-understood, mechanical, repetitive changes with automated checks that would catch the plausible mistakes. - Work that is individually verifiable and cheap to reverse. - Long reading and research passes, where a second person mostly adds interruption. - Anything where one person is genuinely blocked-free and the other would be idle-watching. ## Make it a property of the work, not of the person This is the part that goes wrong most often. If pairing is assigned because of who is doing the task, it reads as supervision, and everyone learns to interpret "you should pair on that" as a judgement. Attach the criteria to the *work item* at planning: this one is hard to reverse, so it is paired; this one touches an area with one knowledgeable person, so it is paired and the newcomer drives. The same rule then applies to the most experienced person on the team, which is exactly what makes it survivable. ## Run it as a bounded default, and revisit Set the pairing default for a named class of change, put it somewhere visible, and re-examine it periodically. Two things should move it: - **Surprises.** If an area keeps producing defects or unpleasant discoveries, it has just declared itself higher-risk than you classified it, and it moves into the paired set. - **Spread knowledge.** Once four people can work confidently in the area that used to have one, the bus-factor reason for pairing there has been discharged and the default can relax. ## Cost and energy are real constraints A paired hour consumes two person-hours, and pairing cannot be sustained at full intensity all day. Selectivity is partly about spending a limited budget where it buys the most, and partly about protecting the practice: a team told to pair on everything will quietly stop, and the sessions that mattered will be lost along with the ones that did not. ## What selection does not buy you Choosing well does not mean a paired change needs nothing else. A pair is two people with a single shared blind spot; automated checks and the pipeline still apply, and so does an independent look from someone who was not in the room, precisely because they do not share the pair's assumptions.