AI pair programming

Pair programming was formalised in Extreme Programming by Kent Beck in the late 1990s: two developers at one workstation, a driver who types and a navigator who reviews each line, thinks ahead and catches mistakes, swapping roles frequently. GitHub marketed Copilot from its 2021 launch as "your AI pair programmer", and tools such as Aider describe themselves the same way. The metaphor is imperfect in an instructive way: in AI pairing the model usually drives, producing code faster than a person types, and the human becomes the permanent navigator, whose job is exactly the review discipline that is easiest to let slip.

Effective practice keeps changes small and reversible. The developer states intent and constraints, asks for one focused change, reads the diff, runs tests, and commits before the next step, so each increment can be rolled back; Aider, for example, commits every AI edit to git automatically for this reason. Useful moves include asking for two alternative approaches before choosing, asking the model to explain a line or predict how code behaves on an edge case, having it write failing tests first and then the implementation, and challenging its assumptions rather than accepting the first answer. The model has no persistent memory of team conventions unless they are supplied through instruction files or the prompt.

The productivity evidence is mixed and context-dependent. In a controlled experiment by Peng et al. (2023), developers with Copilot completed a greenfield task, implementing an HTTP server in JavaScript, 55.8% faster than a control group. METR's randomised trial published in July 2025 found the opposite for experts on familiar ground: 16 experienced open-source maintainers working on 246 real issues in their own large repositories took 19% longer when AI tools were allowed, while believing afterwards that AI had made them about 20% faster. The gap between perceived and measured speed is itself a reason to measure rather than assume.

The security evidence points the same way. Perry et al. (ACM CCS 2023) found that participants with an AI assistant wrote less secure code on several tasks and were more likely to believe their code was secure, a textbook case of automation bias. Human pairing also delivers knowledge sharing, collective code ownership and mentoring, which AI pairing does not replace and may erode if juniors accept code they could not have written. The line from vibe coding is whether the human reads and understands every change; the line from a coding agent is whether the human is present at every step.