Vibe coding forms just one limited aspect of the engineering challenge, according to DoorDash chief executive Tony Xu.

Writing code occupies only a portion of a software engineer's daily work, and modern models help generate it, Xu said on a Tuesday episode of the Uncapped podcast. This covers perhaps 25 percent, 30 percent, or 50 percent of the time spent shipping code.

Xu co-founded the food and grocery delivery firm in 2013. DoorDash went public in 2020 as one of that year's major tech IPOs.

The CEO added that engineers spend considerable time on product reviews, design meetings, and alignment with business teams. These areas also need AI support.

If those elements stay the same, the hoped-for productivity boost may not materialize, he said. Many companies are refining workflows to become AI native not only in code development but across operations.

The firm still needs progress on physical AI uses such as food-preparation robots and autonomous delivery vehicles, Xu noted.

Xu has taken a measured view of AI effects at DoorDash.

During a May earnings call he said AI has not yet altered the company's structure.

We see strong productivity gains from AI, with well over half and likely about two-thirds of code now written by AI, Xu said. Yet that fact alone does not show how workflows and teams should evolve.

He joins other tech leaders who have discussed AI limits.

Earlier this month Wix CEO Avishai Abrahami, whose company bought vibe-coding platform Base44, said the industry overstates AI.

I would like to say our Base44 team reaches a thousand people in three years along with a thousand at Wix and the rest handled by AI agents, but I am not sure it is possible, Abrahami said. We give AI too much credit.

Even Andrej Karpathy, who coined vibe coding and is a former Tesla AI head and OpenAI founding member, has noted drawbacks.

In an April talk Karpathy said AI-written code can be messy.

Sometimes it is concerning because the code is not always excellent, he said. It is bloated with copy-paste, awkward brittle abstractions that work but look poor.