What Survives When an AI Coding Agent Writes the Code

Boris Cherny did not set out to redesign the software profession. He built Claude Code as a side project at Anthropic, the AI lab, and the tool went on to trigger a 2026 market scare and popularize "vibe coding": the practice of telling an AI, in plain language, what you want built and letting it write the code. On Bloomberg's Odd Lots podcast, Cherny argues the larger story is what moves out of the engineer's hands and into their head. Not code. The specifying, reviewing, and judging of it.

Claude Code is Anthropic's agentic coding tool. Where older AI coding assistants suggested the next line or the next function, an agentic tool reads a whole repository, edits files, runs tests, and iterates until the task is finished. Claude Code helped turn that model into a default expectation inside Anthropic and, by Cherny's account, inside a growing slice of the industry. The 2026 market scare the show notes describe came from the same fear the adjacent Bloomberg feature documents: an early case of a white-collar profession where the AI labor shift is observable in the working day, not in a forecast.

That shift reframes what an engineer is for. The skill of writing code by hand, the thing the profession was hired for, is no longer the scarce input. What is scarce is the judgment to specify a problem tightly enough that the agent can act on it, to read what comes back and tell good from bad, and to know which failure modes matter in production. The Bloomberg features piece on the same shift documents the same reframe from the buyer side: engineering managers are beginning to measure review quality and shipped behavior, not lines of code.

"Vibe coding" names the consumer end of that curve. A hobbyist describes an app, the agent builds a prototype, the user ships it without reading the source. Cherny treats the term as useful shorthand for what happens when code-writing falls away as a step the human has to perform. He also flags a gap the same show notes mark: the distance between a weekend vibe-coded prototype and a production system that other engineers have to maintain, debug, and certify. The agent does not close that gap on its own. The engineer's judgment, in his telling, is the part that does.

The conversation does not stay inside the engineering pitch. Cherny and the hosts turn to the safety and alignment work that has to travel with an agent that operates a developer's machine, and to the broader question of who gets access to capable models at all. The source itself notes Jamie Dimon's warning, in a separate Odd Lots appearance, about the consequences of putting these tools into broad circulation before the guardrails catch up. Cherny does not dismiss the concern. The same show notes that describe Claude Code as the hottest piece of software in the world also treat the safety and access questions as load-bearing rather than rhetorical.

One caveat anchors the rest. The full Odd Lots transcript for this episode is not yet in the source packet, so any specific Cherny quote below the show-notes level is hedged here. The shape of the argument, and the profession-shift framing it sits inside, are well supported by the show notes and by the Bloomberg features piece. A reader who wants the line-by-line exchange should wait for the transcript or the Bloomberg audio pickup before treating any one sentence as canonical.

The cleanest test the interview leaves behind is one a team can run on itself. If the group still measures itself by lines of code written or features shipped per sprint, the skillset shift has not yet landed. If the group measures itself by the quality of its specifications, the sharpness of its review, and the behavior of the systems it puts in front of users, then Cherny's reframe is already the working definition of the job.