AI Pair Programming Workflow

A development workflow where the AI agent acts as a pair programmer through three stages: research, parse, and implement. The agent’s output is a rich diff — code plus decision-context comments — that the engineer then edits inline, trimming ephemeral guidance and keeping substantive rationale before accepting.

The three stages

Research — context management. The agent gathers what it needs to understand the problem.

Parse — plan the solution. The agent evaluates multiple approaches with pros and cons. The user makes decisions during this phase. These decisions are the critical human-in-the-loop artifact.

Implement — the agent generates an edit diff. But the diff is different from typical AI-generated code: it includes comments not just about what the code does, but about the decisions behind it. Some comments are ephemeral (scaffolding for the engineer’s review), others are substantive (worth keeping as documentation).

The human’s role

The engineer is expected to edit the diff inline — delete the ephemeral comments, adjust the code, and only then accept it. The agent’s output is a draft for revision, not a finished product. This inverts the typical Copilot/Cursor model where the agent generates and the human accepts or rejects.

Parent discussions

• 01-Log/Daily/Sunday June 28 2026 — sparked on 2026-06-28

Connected concepts

• Human Decision as Development Artifact
• …