The company selling AI agents that can code is too lazy to use them on their own flagship product.
The Summary
- Anthropic's Claude Mac app is still an Electron wrapper 20 months after launch, while ChatGPT shipped a native Mac app that keeps getting better
- Claude's own agentic coding tools can now control your Mac to write code, yet Anthropic won't use them to rebuild their slow, buggy desktop experience
- The gap between what Anthropic sells and what Anthropic ships reveals the actual state of AI coding agents: good enough for demos, not yet trusted for production
The Signal
Anthropic has been running a masterclass in unintentional honesty. Since October 2024, their Claude Mac app has been a performance art piece demonstrating the limits of what they actually believe their own technology can do. The app is Electron, that web-wrapper framework that gives you cross-platform reach and users a slow, memory-hungry mess that feels alien on every operating system it touches.
Meanwhile, OpenAI shipped ChatGPT as a real Mac app. Native Swift. Feels like it belongs on the platform. Gets better with each update. The kind of thing you don't think about because it just works.
"If Claude Code is so good, why don't they use it to make an even halfway decent native Mac app?"
Here's what makes this more than just Mac nerd drama. Anthropic now sells Claude Code and Claude Cowork, agentic tools that can take control of your computer to write software. They demo these capabilities in slick videos. They price them for enterprise buyers betting their roadmap on AI-assisted development. The pitch is clear: agents can handle the grunt work of turning specs into shipping code.
But when it came time to rebuild their own desktop app, the one their users interact with daily, the one that represents their brand every time someone switches from their browser? They stuck with the Electron clunker. For 20 months and counting.
Key tells:
- They trust their agents enough to market them
- They trust their agents enough to charge for them
- They don't trust their agents enough to use them for their own product
Drew Breunig nailed the broader implication back in February. If agentic coding actually works at the level being sold, the Electron era should be over. You write one spec, one test suite, and let agents generate native code for each platform. Small teams ship great experiences everywhere. Users get fast, platform-native apps instead of web wrappers pretending to be software.
That's the theory. Anthropic's shipping behavior tells you the reality. Either their agents can't reliably rebuild a desktop app to native code quality, or the internal friction and risk of betting on agent-generated production code is still too high for their own engineering team to stomach.
The Implication
This is your bellwether for where agentic coding actually is versus where the demos say it is. When the companies building and selling these tools start using them for their own production applications, not just internal prototypes or marketing videos, that's the signal the technology has crossed the chasm.
Until then, treat agent-coding claims like you treat any vendor pitch: trust what they ship to their own users, not what they promise you can ship to yours. The Anthropic team knows something about their agents' capabilities that they're not putting in the keynote slides. They're showing it to you every time that Electron app beaches the RAM ball.