TL;DR: AI pair programming is coding with an assistant in a tight loop - it completes lines, explains unfamiliar code, drafts functions, and suggests fixes while you type, and you accept or reject every change as it happens. You stay the driver; the AI is a navigator that never gets tired and never gets offended.
How it works
The name borrows from extreme programming, where two humans share one keyboard: a driver who types and a navigator who watches, questions, and catches mistakes. AI pair programming keeps the roles but swaps the navigator for a model that sees your open files and recent edits. It shows up in three forms: inline completions that predict your next lines as you type, a chat panel that answers questions about the code on screen, and edit commands like "extract this into a helper and add error handling" applied to a selection.
A concrete session: you are adding a date filter to a reporting page in a codebase you barely know. The assistant autocompletes the query change from the pattern in neighboring code, you ask the chat panel why the existing tests stub the clock, it explains the timezone edge case, and you have it draft the test for the new filter - reading and adjusting each suggestion before accepting. Nothing lands without your eyes on it. That constant review is the defining trait.
That is also what separates pairing from delegation. Hand a whole ticket to an AI coding agent and you review the result at the end; pair with an assistant and you review continuously, keeping a full mental model of the code as it forms. Pairing trades some speed for that understanding, which makes it the right mode for unfamiliar codebases, subtle changes, and learning. It is effectively the strictest form of human-in-the-loop AI: the loop closes on every suggestion.
The main pitfall is automation complacency - tab-tab-tabbing through plausible completions until a wrong one slips by. Suggestions are pattern-matched from context, so they confidently reproduce whatever the surrounding code does, including its bugs. Skilled pairs read suggestions with the same suspicion they would give a human partner's idea, and drop to unassisted typing when precision matters more than speed. In practice most developers slide along a spectrum all day, pairing on delicate work and delegating the rote parts - the skill is knowing which mode the moment calls for, and both modes reward the same clear intent that vibe coding runs on entirely.
Where it sits in the AI stack
Pairing lives in the editor, between the developer's intent and the codebase, with the human approving every step:
Key tools and implementations
-
GitHub Copilot
The tool that made inline AI completion mainstream, embedded in most major editors.
-
Cursor
An editor built around AI pairing, from multi-line tab completion to chat grounded in your open code.
-
JetBrains AI Assistant
Completion, chat, and refactoring help integrated across the JetBrains IDE family.
-
Terminal assistants
Tools like Claude Code used conversationally - a pairing partner for developers who live in the shell.
Related entries
- AI coding agent Software that plans and executes multi-step coding tasks - reading files, editing code, and running tests with minimal supervision.
- Vibe coding Building software by describing intent to an AI and accepting code you judge by behavior rather than by reading it line by line.
- Context management Keeping an AI agent's working context relevant and small through compaction, memory files, and delegating work to subagents.
- Human in the loop A design pattern where a person reviews, approves, or corrects an AI system's actions at defined checkpoints.