When your entire curriculum becomes obsolete in one semester, you either redesign the course or admit college is just expensive babysitting.

The Summary

  • Stanford's CS146S redesigned its entire AI engineering curriculum because coding agents can now complete last year's assignments in a single prompt
  • Students now work on real open-source repositories with actual maintainers, proposing changes to unfamiliar codebases instead of synthetic homework problems
  • Mihail Eric from Monaco AI runs the course, bringing industry directly into the classroom while the traditional assignment-test-grade model crumbles

The Signal

Stanford just ran the second iteration of CS146S, and the course design tells you everything about where software education is headed. The core problem: what do you assign when a coding agent can finish last year's homework in one prompt? Mihail Eric's answer is to abandon synthetic exercises entirely and make students do what actual engineers do — dive into someone else's codebase, understand it, and propose a change worth merging.

The course structure flips the traditional model. Students pick an open-source project they've never seen before. They learn the repo's architecture, idioms, and contribution guidelines. Then they propose a real feature or fix, defend it to the maintainer, and ship it. No contrived problem sets. No auto-grader. Just the work.

"Students have to learn an unfamiliar repo, propose a change, and get it merged by actual open-source maintainers."

This is AI-native education. Not "here's how to use ChatGPT" or "prompt engineering 101." It's recognizing that the value of a programmer has shifted from writing code to understanding systems and knowing what to build. When synthesis is free, taste and context become the scarce resource. The course teaches that.

Monaco's involvement here matters because Eric is running a company built on the premise that agents will write most of the code, which means he's teaching students to work *with* agents, not compete with them. The curriculum assumes agents as infrastructure, the same way web development courses assume HTTP exists.

Key shifts in the course model:

  • Open-source maintainers replace TAs for project feedback
  • Understanding existing systems replaces building from scratch
  • Real-world code review replaces synthetic grading rubrics

The second run of the course means this isn't an experiment anymore. It's a template. Stanford is iterating on a model where students produce actual value instead of homework that gets thrown away. The externality is that open-source projects get contributions from smart undergrads. The learning outcome is that students experience the messy reality of production code, not the clean fantasy of leetcode.

The Implication

If you're hiring junior engineers, start asking what repos they've contributed to, not what their GPA was. The students coming out of courses like CS146S have built context-reading and system-understanding muscles that traditional CS grads never needed. They've also learned to use agents as leverage, which means their output ceiling is higher.

For educators, the CS146S model is the blueprint. Synthetic assignments are dead weight now. Real projects with real stakeholders force students to develop judgment, which is the only durable skill when agents can write the code.

Sources

Latent Space