State of Agentic Coding #7 with Armin and Ben

TL;DR

• 1

  Most serious users still are not "Ralph looping" 24/7: Despite hype from figures like Anthropic's Boris Cherny and OpenClaud's Peter Steinberger, both Armin and Ben say they still work mostly in the terminal with close supervision because they do not know how to QA or even comprehend code produced by fully autonomous loops.

• 2

  Token spend has become the clearest revenue metric the software industry has ever had: They argue model providers love features that increase token usage because, unlike DAU or impressions, more tokens map almost directly to more dollars, which explains why Claude, Copilot, and others are tightening subsidies and pushing usage-based behavior.

• 3

  AI is creating a security arms race that forces companies to buy AI either way: Ben and Armin describe a bizarre loop where AI tools find and file more security issues, which then forces maintainers and companies to spend tokens on AI-based security review just to keep up.

• 4

  The Bun rewrite from Zig to Rust is a real test of AI-driven code migration: After Bun creator Jared Sumner's company was acquired by Anthropic, the project was largely rewritten from Zig to Rust with agents, and Armin says if it stays stable, enterprises will treat it as evidence that legacy ports are now practical.

• 5

  Languages optimized for human creativity may lose ground to machine-friendly ones: Armin argues Ruby and Zig make trade-offs that humans love, such as metaprogramming or easy cross-compiling, but those same traits make them worse for LLMs, raising the unsettling prospect that AI coding favors languages like Rust over more expressive human-first options.

• 6

  Local open-weight coding models are finally becoming usable, but they feel like Claude from a year ago: Armin says projects like Andy Rees's DwarfStar 4 make 96 GB-class local models practical enough for real coding work on high-end Macs, yet the experience still feels noticeably behind frontier hosted models.