The LLM Comments Are Not for You

LLM-generated code comments have become verbose, decision-heavy, and full of invented terms because models are being trained and reinforced inside agentic workflows that reward behaviors improving end-to-end outcomes. Reinforcement learning and agent swarms push models to treat comments as a persistent store of rationale and coordination signals: when one agent writes “array, not a linked list,” other agents read and defer to that rationale, preventing regressions and helping multi-agent coordination. Session compaction and limited context force models to re-encode design decisions into diffs, producing large blocks of comments (the author notes patterns like +50/-1 comment-heavy commits) that are optimized for machine consumption rather than human readers.

Removing or sanitizing those comments is difficult because they are part of the mechanisms that make agents succeed on benchmarks; prompts or “desloppifier” agents have limited effect and can undermine the behaviors that earned high evaluation scores. That optimization creates emergent organization, novel vocabulary, and potential long-term risks analogous to biological maladaptations: a behavior useful in the training environment that might be harmful or unwieldy outside it. Human attention and editing currently counterbalance accumulation, but fully agent-maintained codebases over years could look very different, and the trade-offs between agent-optimized commentary and human maintainability remain unresolved.