An audit that keeps you in the room

A piece of news worth looking at made the rounds these past weeks: Bun, a widely used tool from the JavaScript world, was rewritten from one language into another on the back of dozens of AI agents — some writing the code, and others, in parallel, hunting for flaws in the first ones’ work, like adversarial reviewers. A genuine milestone: one person orchestrating a job that by hand would have taken a team close to a year.

We care about it for what it says about audits in the age of AI. And there are two truths worth holding at once.

Each tool to its task

The first: for a job like Bun’s, having AI audit AI at scale is exactly right. It was, at bottom, a mechanical port — translating a program from one language to another while keeping the same behavior. There, “correct” has a clean definition: the tests pass, the behavior is the same. It’s a target a machine can check a thousand times without tiring, and setting AI reviewers to hunt for bugs is a brilliant way to scale that checking far beyond what a human would read line by line.

But most engineering is not a port. Most of the time the right answer isn’t a green test but a judgment: is this the right design? Does it solve the real problem or a more convenient one? Where should this go next? There’s no target the machine can mark on its own, because the target itself is a decision.

Where the human sits

And here’s the nuance, because it’s easy to get wrong. Even in Bun’s story the human doesn’t disappear: the person who directed it spent days watching the process, fixing the machinery when it broke, and pressing the final button himself. The human didn’t leave — he sat down to tend the machine.

In StrayMark, audits are also run by AI agents, but their value rests on where we put the human: not tending the machine, but paired with the substance. The person directs who audits, decides, and adjudicates when several independent model families agree — because that agreement, and not the word of a single model, is the signal. We recently reinforced exactly this: we made the identity the operator declares override what the tool believes about itself. The audit stopped trusting its own guess about who ran it, and deferred to the person. We tell the full story in Who the audit thought it was.

The value of an audit isn’t in taking you out of the room. It’s in letting you stay in it without having to read every line — directing, deciding, trusting what you see.

Not on the sidelines, but in command

These are two bets on AI that don’t compete: Bun’s shines when the task has a checkable answer; ours guards the far more common case where the answer is a judgment. Both have a human — the difference is what we ask of them. We don’t ask them to watch the machine from the outside, but to stay paired with it, oriented and in command. It’s the same old conviction, and we keep building it in StrayMark.