Stanford rewrote CS146S for agents: open-source patches replace one-shot homework

Latent Space · rss · 2026-11-09

Latent Space interviews Mihail Eric — Stanford lecturer for CS146S (The Modern Software Developer) and head of AI at Monaco — on teaching and practicing AI-native engineering.

Course redesign

Last year's 9 assignments + capstone got rewritten: every surveyed Stanford sophomore/junior already codes with agents, so much of the old homework is now one-shot-able.
The new core assignment: contribute a patch to a real open-source project, reviewed by the maintainer — forcing students to learn what the maintainer wants, why the code works that way, and what the patch might break.
He recommends Geoffrey Litt's explain-diff skill: turn a diff into a visual walkthrough and quiz yourself on it to verify you understand agent-written code.

What Monaco buys vs. builds

Bought code review tooling and kept Devin; built agent workspaces in-house because they're coupled to Monaco's cloud. Rule: "You buy to learn, and you build to keep."
Built Hal: engineers use it to investigate outages and open PRs; salespeople use it to explain the product — one system, two useful output types.
Cites a Faros study: high-AI-adoption teams merged 98% more PRs but review time grew 91% — code review is the new bottleneck.
Prediction: open-source repos will become agent-run software factories, though humans still decide which changes belong.

2-hour video with full timestamps and transcript.

Original post →

Stanford rewrote CS146S for agents: open-source patches replace one-shot homework

More from coding & agent