Erik Bernhardsson, co-founder and CEO of Modal, estimated that "probably" 95% of the lines of code merged at his company are written by AI agents "to some extent." He gave the figure on the Unsupervised Learning podcast, in an episode published October 28, 2026. Patrick Chase, an investor at Redpoint Ventures who led Redpoint's investment in Modal, had asked how building software at the company had changed.
Modal builds cloud infrastructure that customers use to run AI applications. Bernhardsson said it now powers many AI applications he had no idea would exist when the company started four or five years ago. The publisher's episode notes say Modal has passed a $1 billion run rate. Chase said Modal's roadmap at each board meeting is "extensive and very daunting," and that he is impressed by how much the team ships while keeping the system running.
The agents in question are coding agents: AI systems that write code, run it, read the errors and try again over many steps, with less line-by-line direction than an autocomplete tool. Bernhardsson's 95% was a rough estimate, hedged with "probably" and "to some extent."
Still under close supervision
Bernhardsson said he had spent the past few months writing "an insane amount of code using agents" himself. "It's kind of incredible like what they can build," he said, but humans still need to supervise them "quite a lot."
In his account, infrastructure is what makes that supervision necessary. It carries tail risks, the rare but costly failures around reliability and making sure systems keep working as they scale. "At least for now, I think software engineers are pretty safe," he said. Modal still benefits from having very good engineers to supervise the agents, he added, "probably for the next few years." He came back to the point a moment later and hedged it again: humans seem safe as software engineers for the next few years, he said, then added, "I don't know. We'll see."
Senior engineers use agents most
The adoption pattern surprised him. "I thought it would be the junior people that would adopt agents the most," Bernhardsson said. "I've actually seen almost the other way around."
His explanation is that agents give very senior engineers something like "10 interns." Those engineers have long been held back by how fast they could type, he said, while carrying many ideas in their heads. With agents they can get those ideas out much faster.
He described the kind of task where this works. Agents are good at building against a test suite, so an engineer can hand one an extensive set of tests implementing POSIX, the long-standing standard that defines how Unix-style file systems should behave, and tell it to write a new file system. In his telling, the agent simply goes off and writes one, taking about three days.
Some very senior engineers, he said, have gone "complete agent maxing." "You can move very fast with them," he said.
More code puts pressure on the pipeline
Chase asked whether that pace had changed how Modal handles code review, deployment and CI/CD. CI/CD stands for continuous integration and continuous deployment, the automated systems that test every change and push approved code into production.
"A little bit," Bernhardsson said. Modal is "coming up against some of those constraints" and has had to spend a lot more time on CI/CD. His own view hasn't changed much, he said: long before agents, during his many years as a chief technology officer, he was already obsessed with automated deployments and automated tests. What has changed is that agents are now pretty good at writing tests, which they weren't a year earlier. He did not think Modal had yet run into the foundational limits. In an earlier conversation on Latent Space, he described a related choice: Modal treats code review as optional and leans on automated testing instead.
For a picture of those limits, he pointed to Anthropic. Bernhardsson said the AI company had announced "a couple of days ago" that it was moving off Git, "or at least GitHub." Git is the widely used version-control system that records every change to a codebase; GitHub is a popular service built around it. The likely reference is Anthropic's October 13, 2026 post on redesigning source control, which describes replacing Git in its internal monorepo, a single repository holding much of a company's code, with a system called Scion.
According to that post, daily merged changes at Anthropic grew from about 2,500 in May to more than 30,000 in September. Seven engineers built Scion and related tools and moved the repository off Git in five months, and Anthropic reports a 200-fold cut in the processing power and memory its source control needs compared with its previous Git mirrors. The company also shifted more testing to after code is merged and before it is deployed, because a single, globally locked merge queue could not keep up with its workload. Engineers kept responsibility for architectural constraints and for the intent behind reviews. Anthropic presents Scion as specialized infrastructure for unusually large, agent-heavy repositories.
Bernhardsson expects Modal could face the same problem later. "We might run up against those limits at some point," he said, when the company has 10 times as many commits every day.
Hiring has barely changed
Later in the episode, Chase asked whether coding agents had changed how Bernhardsson thinks about hiring. "Not much," he said. "I think we're still going to need humans for a long time."
His reasoning is historical. Over the last 30 to 50 years of software engineering, he said, every tool that made engineers more productive, from compilers to high-level programming languages to CI/CD, pushed demand for them up. He expects the same now, and said the industry will need many more engineers.
He expects one group to have it harder for a while. In the short term, he said, things may be "a little bit harder" for new graduates, except those at the very top of the range. In the long run he is "very bullish" on computer science. "Ironically, I didn't study computer science," he said, but if his kids wanted to, "I would definitely encourage them."