Counterpoint: deepswe eval found banning comments actually improves agent performance
nateberkopec · x · 2026-10-14
nateberkopec pushed back on treating comments as token-saving discovery tools for agents, listing failure modes:
- Token cost of keeping comments correct
- Wrong comments causing more token waste
- Old comments from dumber agents misleading newer, smarter agents
He argues the question should be settled by evals, noting deepswe banned comments and agents only performed better. He agrees connecting code to human "why" matters but says comments are the wrong tool in 2026.
More from coding & agent
- JevGPT: dual-system autonomous research agent matches rivals faster and cheaper — Yasheng Sun · 2026-10-14
- Dev builds Grok Bot chaining Suno, Midjourney and Opus to autonomously make music videos — TinfoilTricorn · 2026-10-14
- Musk reshares the two-agent setup: Grok for work, Muse for personal life — elonmusk · 2026-10-14
- Dev's one-week finding: Three.js + low-poly is the fastest path to satisfying 3D UIs — dxbhiro · 2026-10-14
- Coatue AI Head Frames Investing as a Compression Problem Solved with Agents — AccBalanced · 2026-10-14
- Emil Kowalski: Outsourcing Everything to Agents Turns Creators Into Efficient Producers — itsOmSarraf_ · 2026-10-14