Stanford is now formally teaching AI-native software development. And in its second year, the course has moved from showing students the tools to drilling them on core principles. Latent Space reports that Mihail Eric, an adjunct professor at Stanford and head of AI at the startup Monaco, has reworked CS146S, “The Modern Software Developer.” The course is getting a lot of attention because it collects current best practices for building software with coding agents into one curriculum.
What stands out is the shift Eric describes. Last year’s version was a broad tour. This year’s version treats a few principles as settled.
Why does a course like this exist at all?
Eric says the idea came to him in November 2024. Coding agents were still early, but he could already see that “something drastically changing” was happening in how developers worked and how fast they shipped.
The trouble was that the know-how was spread across scattered blog posts. Nobody had organized it. So Eric set out to turn it into a formal curriculum. He took inspiration from his friend Chip Huyen, who had built a machine learning systems design course at Stanford. Getting the course through Stanford’s committee took about nine to ten months. It first ran in fall 2025, before the newest generation of frontier models arrived. Students liked it, and Business Insider wrote about it.
What changed between year one and year two?
A lot. In 2025 the market was splitting into point solutions, according to Eric. Separate startups were building coding agents for code review, for spec design and for SRE work (site reliability engineering, the people who keep production systems running).
So the first version of the course was mostly about exposure. Eric calls it a kind of “odyssey” through how every stage of the software development lifecycle was changing.
This year, autonomous coding agents “really picked up in a way that we had not seen before,” he says. The material now centers on principles that are holding up in practice:
- Careful context engineering. That means choosing exactly what information the agent sees, and when, so it doesn’t drift or hallucinate.
- Design and architecture as the main skill. Hard tasks get handled by thinking through the system’s structure before any code gets generated.
That tracks with the wider industry. Once agents can write most of the code, the leverage moves up a level to specs, structure and the context you feed the model.
Is this replacing traditional computer science?
No, and Eric is clear about that. The course isn’t meant for beginner programmers. Classic courses like algorithms are prerequisites.
That matters. A lot of the public debate pits “learn to code” against “let AI code.” Stanford’s answer is to do both, in order: learn the fundamentals first, then learn how to direct agents. If you don’t understand the system, you can’t judge what the agent hands back.
Who feels this shift first?
New graduates. In Eric’s words: “This particular generation of junior developers, of junior engineers, the people that are now entering the workforce, will, I think, be the first kind of generation of that new shift.”
That’s a big claim, and there’s reason to take it seriously. Eric isn’t only lecturing. He runs AI at a company that ships with these tools every day. That practitioner view gives his syllabus more weight than a purely academic take would have.
What should teams do with this?
If you hire or manage engineers, a few takeaways:
- Rethink onboarding. New hires may show up fluent in agent workflows. Make sure your codebase and docs actually support that, with clear specs and context files agents can read.
- Treat context as infrastructure. Prompts, rules files and architecture docs now affect output quality as much as the code does.
- Test for judgment as well as typing. Interviews should check whether candidates can break down a problem and review AI output critically.
- Expect the tooling to consolidate. The narrow tools for each stage of development are giving way to general agents. Don’t over-invest in niche tools.
What comes next?
The interview aired as the AI Engineer YouTube channel passed 1 million subscribers, and just before the AI Engineer CODE Summit. Host swyx listed the summit’s themes as fighting AI “slop,” fast models, and the confidence to attempt major rewrites. Universities are now writing the agent era into their official course catalogs, so the best practices are settling into a standard. Developers who learn them early will have a real advantage.
The full conversation, including Eric’s topic-by-topic walk through the syllabus, is available on Latent Space.