HN: Opus 4.5 is not the normal AI agent experience that I have had thus far

Hacker News discussion (879 points, 1,353 comments) on a blog post by tbassetto describing their experience building applications with Claude Code + Opus 4.5. The OP created multiple apps (image converter, SVG editor, etc.) without understanding how they were assembled, claiming Opus 4.5 represents a qualitatively different tier of coding agent.

The thread is a Rorschach test for the state of AI-assisted development in early 2026. Comments range from "I don't write Rust code anymore, I just read it" (rtfeldman, Zed) to "on novel work it's a hindrance" (JDye, proxy servers) to "it's a better programmer than I am, it's just not anywhere near as good a software developer" (jaggederest).

Key themes that emerge:

Compiler-as-guardrail: Strongly typed languages (Rust, C#) give agents faster feedback loops via compile errors. Multiple commenters report better results with Rust than Python/JS because the compiler catches mistakes instantly. gck1 switched to Rust for everything, even throwaway scripts. "javascript can still have ugly runtime WTFs" (tezza).

Training-data proximity: LLMs excel at well-trodden patterns (React/TypeScript) but struggle with novel or low-training-data domains. C++ 3D graphics, MASQUE proxying protocols, and bleeding-edge Swift all produce worse results. dpc_01234: LLMs are "fundamentally... non-deterministic pattern generators" — renaming confusingly similar things in a codebase fixes performance.

Skeptic-conversion workflow: kevin42 describes converting skeptics with 15-30 minutes of prep: document architecture, create subagents per component, use planning mode. Everyone responds positively. The thread is full of former skeptics (ryandrake) who changed their mind with Opus 4.5 specifically.

Programmer vs. engineer: parliament32 gatekeeps "engineering" for novel work, drawing on Goedecke's pure/impure engineering distinction. Counterpoint: most software engineering applies known tools to non-groundbreaking problems (emodendroket).

Bespoke software thesis: ryandrake argues the 10% usage adage — people only use 10% of bloated software features — can now be solved by each user building their 10% for themselves. apitman: AI-built "bad code" that's smaller than bloated alternatives with tailored UX could be compelling.

The thread was partially truncated during fetch (~135 of 1,353 comments captured).