LLM Coding Agents Have a Serious Quality Ceiling

Multiple HN threads are converging on the same uncomfortable finding: LLM coding agents are genuinely useful but break down in specific, predictable ways. A paper on 'constraint decay' found that LLMs perform well at open-ended code generation but fall apart when forced to navigate explicit architectural constraints. A separate thread titled 'Claude is not your architect' made the same point from lived experience: left to its own devices, Claude will confidently design systems that don't hold up to scrutiny, and it will defend those decisions even when wrong. The DeepSeek-native coding agent thread added another dimension, with developers shopping around for the cheapest capable agent rather than the best one.

The pattern here is that developers are starting to map the actual capability frontier rather than just celebrating it. The enthusiasm of 'agentic coding changed everything' is giving way to 'here is exactly where it breaks.' Constraint-following, long-horizon consistency, and architectural judgment are the weak spots, and they are the spots that matter most for production software.

The counterpoint, noted in the threads, is that models keep improving and some users report a step-change in capability in the last year. But the more technically sophisticated commenters, the ones who know what good architecture looks like, are the most skeptical. That gap between what AI can generate and what experienced engineers would approve is the real signal.