# Review Pipeline: How to Review Code When AI Wrote 80% of It
Traditional code review assumes code is human-written and every diff can be read line by line. The AI4SE era needs a new review paradigm — shifting from line-by-line inspection to spec-conformance acceptance.
The Old Paradigm Breaks Down
Traditional code review has implicit assumptions:
- Change volume is small (a few hundred lines)
- Humans understand the intent behind every line
- Reviewer and author share common context
AI-generated PRs can be:
- Thousands of lines / dozens of files
- Generation logic humans don’t understand
- Agents with no “original intent” you can ask about
New Paradigm: Three-Step Review
| Step | What to do | Who does it |
|---|---|---|
| Spec conformance | Diff vs. Spec — did it produce what it should have? | Human + Agent |
| Quality scan | Automated checks for security, performance, maintainability | Agent (independent Evaluator) |
| Human judgment | Is the architecture sound, are the trade-offs acceptable? | Human |
Practical Recommendations
- Agent generates + another Agent reviews ≠ safe
- There must be an Evaluator independent of the generation chain
- PR templates should include a Spec link and acceptance criteria
- AI PRs over 500 lines should be split into smaller submissions