[HN]score: 0.37
Why LLM Agents Fail at Modifying Real Codebases Beyond Simple Demos
May 27, 2026
Current LLMs can handle additive tasks like reading and planning in codebases but cannot safely perform transformative modifications—editing dependencies, invariants, and causal structure—in real-world repositories beyond toy examples of tens of lines.
HOW THIS AFFECTS YOU
●
builderYou should scope AI coding agent features to additive tasks (summarization, planning, greenfield generation) rather than autonomous refactoring of production codebases.
●
researcherThe additive vs. transformative distinction is a useful framing for evaluating where current agent architectures actually break down in software engineering benchmarks.
●
founderProducts promising autonomous end-to-end software delivery on real repos are overstated; the defensible near-term market is augmentation, not replacement.