AI coding agents patch single bugs instead of rethinking design, engineer warns
neal_lathia · x · 2026-11-13
Engineer Vlad Tokarev argues AI coding agents are poor at stepping back: a human engineer hitting a bug can question earlier design decisions and redesign so a whole class of problems disappears, while agents patch the single case.
Prompting doesn't fix it — even when explicitly told to ignore existing patterns, agents stay tied to them. A popular comment suggests a workaround: add a rule to AGENTS.md to file design tensions, bugs and usability issues to GitHub with context, then run a dedicated issue-sweep agent that looks for themes and considers redesigns. Another commenter notes LLMs are inherently reluctant to delete lines of code.
More from coding & agent
- Arena's Interaction-as-a-Judge: evaluating computer-use agents by how they play — arena · 2026-11-13
- Current launches as an opinionated tool to run hundreds of AI agents in control — edgarpavlovsky · 2026-11-13
- HermesTube plugin brings YouTube subscriptions into Hermes agent, skipping the algorithm — Teknium · 2026-11-13
- RepoMem: Giving Coding Agents Long-Term Memory from Git History to Improve Bug Localization — burkov · 2026-11-13
- The Zvi: Opus 5.5 keeps finishing my coding ambitions before I can use Fable 5.2 — TheZvi · 2026-11-13
- Mnemos gives Claude Code, Codex and Hermes one shared cross-agent memory — RileyRalmuto · 2026-11-13