HN post catalogs five failure modes of AI coding assistants

A developer posting as sollawen laid out a specific list of complaints about agentic coding tools that pulled 33 points and 22 comments on Hacker News. The five gripes are worth naming in the poster's own words because they line up with what a lot of practitioners have been saying privately: reinventing existing code inside the same file, ignoring effects on the rest of the system, blowing through the 200k token limit, degrading once the context is packed, and reaching for new architecture when a small change would do.

The phrasing is specific. The poster calls out AI assistants for "writing three duplicate functions for the exact same feature in a single file" and having "zero holistic awareness" as they hyper-focus on the current task. On long context, they described the behavior as "instant brain damage" once the window is full. On logic, they noted the model will "insist on designing an entirely new system architecture" for a change a human developer would knock out in three to five lines. The models they cite by name are GLM5 and miniMax-3.

The comment thread does not agree on a diagnosis. A commenter posting as coldtea asked whether the poster had actually directed the model to consolidate code and reuse existing implementations. Others treated the problems as real but tractable through process: plan first in markdown, add an AGENTS.md or CLAUDE.md that spells out architectural standards, split work across specialist agents for requirements, architecture, coding, review and testing, and move to models advertising roughly one-million-token contexts. One reply framed the mindset shift as treating AI as "a knowledgable, patient tutor rather than let it code for me."

The honest caveat is that this is one developer's frustration post, not a benchmark, and the models named are not the flagship coding assistants most teams reach for by default. What the thread does not give you is any measurement of whether things have actually gotten worse over time, or whether the reported failures are model-side, tool-scaffolding-side, or workflow-side.

The reason the post is worth reading anyway is that the failure modes it names are the exact seams where the current agent-coding stack is thinnest, and the top replies cluster around the same fix: put more human structure around the model rather than ask it to hold the whole codebase in its head.