bun just got rewritten in rust — by an agent swarm, in 11 days

for two years, the standard skeptic line about vibe coding went like this: “sure, it can scaffold a todo app. wake me up when it can do a realcodebase.”

consider this your wake-up call. according to reporting on anthropic's may 28 update to claude opus 4.8, bun creator jarred sumner demonstrated the new dynamic workflows feature by pointing it at one of the gnarliest jobs in open source: migrating the bun javascript runtime from zig to rust. the reported numbers: roughly 750,000 lines of code, 11 days, and a 99.8% test suite pass rate at the end.

dynamic workflows, if you haven't touched them yet, are the headline of the 4.8 update: instead of you hand-wiring a pipeline of subagents, the model writes its own orchestration script — deciding what fans out, what verifies, and what synthesizes — then runs a crew of subagents against the plan. it's the difference between hiring one very fast intern and hiring a very fast intern who is allowed to hire.

the asterisks, because we read the fine print

a demo from the person who knows the codebase best is not the same as your monday morning. sumner is not a typical operator — he's spent years inside bun's internals, and a migration is exactly the kind of task agents are best at: the destination is known, the behavior is pinned, and the work is wide rather than deep. 99.8% passing also means thousands of tests not passing on a suite that size, and the last 0.2% of a runtime is where the dragons live.

so no, this doesn't mean “agents can do anything now.” it means something more specific and honestly more useful: agents can now hold a workload that used to require a team, for weeks, without losing the plot — when the rails are good.

the real hero is the test suite

here's the part the hype posts keep skipping. the reason this migration was even attemptable is that bun has a brutal, comprehensive test suite. the tests were the contract. the agents weren't trusted — they were verified, hundreds of thousands of times, automatically.

the lesson for vibe coders isn't “buy more agents.” it's “write the checks that make agents safe to unleash.”

that's a thing you can do today, on your own project, at any size. every test you write is permission for an agent to move faster later. the builders who internalize that are going to spend 2026 shipping migrations and rewrites the rest of us keep postponing.

// if you want the technique version of this take — our guide on the four prompts to run before any feature is basically this article in checklist form.