Line-by-line code review can't scale with AI agents, so these devs went function-level
Wise_Reflection_8340 · reddit · 2026-10-05
The authors found agents now produce more code in a week than a team used to in a month, making line-diff review unscalable: real logic changes hide among renames, moves, and reformatting.
Their approach: review at function granularity instead.
- Mark each function as changed, renamed, moved, or only reformatted—cosmetic ones can be confidently skipped;
- For actually changed functions, look at dependents—a small edit called from 40 places deserves far more attention than a big edit to a leaf function;
- Call graphs also expose architectural drift when agents quietly add dependencies across boundaries.
They open-sourced two tools: sem shows function-level diffs and call relationships; inspect ranks changed functions by potential blast radius so reviewers (human or model) start with the riskiest few.
More from coding & agent
- SuperX MCP in Claude Code grew an X account from 64 to 1,000+ followers since July — tibo_maker · 2026-10-05
- Vanishing Gradients podcast: agents are like coffee — they help you make stupid mistakes faster — hugobowne · 2026-10-05
- Codex resets only unlock near your limit — tip: burn a bit of usage first to redeem expiring resets — JeremyNguyenPhD · 2026-10-05
- A permit layer for MCP tool calls makes same-key retries return the original receipt — HotPocketWaves · 2026-10-05
- AI agent produces a 1:45 video end-to-end on a local PC with Pixio and FFmpeg — tsi_org · 2026-10-05
- Stripe and Tempo Launch MPP, a New Payment Protocol Built for AI Agents — glenbeer · 2026-10-05