Last weekend, developer and AI observer Simon Willison published a post on Hacker News noting that Anthropic had quietly shifted Claude Code — its command-line AI coding assistant — to run on Bun, a JavaScript runtime itself written in Rust. The change wasn't announced with fanfare. It appeared in the tooling, and the community noticed.
That is not a crisis. It is a signal.
What's actually changing
The specific technology here — Bun, Rust, runtime environments — matters less than the pattern it represents. AI tools are not maturing into stable utilities. They are still in a phase of rapid, sometimes invisible architectural change. The underlying engine of a tool you used last month may be something entirely different this month, without your knowing.
For developers this is par for the course. For everyone else — the small-business owner using Claude to draft contracts, the parent using an AI assistant to help a kid understand a homework concept, the household budget manager using AI-generated spreadsheet formulas — it raises a question worth sitting with: how dependent are you on tools that may change, break, or disappear without notice?
This is distinct from the standard advice to "not trust AI." The more pointed issue is vendor lock-in at the workflow level. When a tool rewrites its internals, behavior can shift subtly. Outputs that were reliable may change. Integrations can break. A small business that automated its invoice processing around a specific AI behavior may find, after an invisible update, that the output format changed.
The AI industry has no equivalent of the pharmaceutical industry's change-notification requirements, or the food industry's labeling standards. Tools update when the company decides they should.
What we'd actually do
Audit which AI tools you actually depend on, not just use. Dependency means your workflow breaks without it. Usage means it's convenient. Write that list down — it takes ten minutes and most people are surprised by how long it gets. Dependency items warrant more scrutiny than usage items.
If you're running a microbusiness and an AI tool is handling any step in a revenue-generating workflow, you need a manual fallback for that step. Not because the tool will fail this week, but because it will eventually behave differently than it did when you built the process around it. The fallback doesn't have to be fast. It has to exist.
For AI-assisted work you want to preserve, export and version your outputs, not your prompts. Prompts are instructions to a system that changes. The useful artifact is the output. If Claude Code, or any AI assistant, is helping you build something — a template, a script, a formatted document — save the finished output in a format you control, not inside the platform. Local files, not cloud-native. Text and PDF, not proprietary formats.
Treat AI tool updates the way you treat software updates generally: don't be on the bleeding edge for critical tasks. If you use an AI coding assistant or productivity tool professionally, consider maintaining a known-working version or configuration and only migrating to updates after a week or two of community feedback. This is basic IT hygiene, and it applies to AI tools now.
Learn one durable skill that underlies whatever the AI is doing for you. If you use AI to write formulas, understand basic spreadsheet logic. If you use AI to draft legal language, understand the structure of what you're signing. The AI handles speed; you supply judgment. That division only works if you have something to supply.
The bigger picture
The Bun-and-Rust rewrite of Claude Code is genuinely unimportant on its own. It may make the tool faster and more stable. The engineers at Anthropic almost certainly have good reasons for it.
What it illustrates is that the AI layer of household and small-business infrastructure is still being built in real time, by companies with their own engineering priorities, on timelines that have nothing to do with your workflow. That's not a reason to avoid the tools. It is a reason to hold them lightly — to keep the useful outputs, preserve the manual skills, and avoid building processes that have a single AI tool as a load-bearing wall.
Durability isn't about stockpiling for the worst case. It's about not being surprised when ordinary things change without warning.
