Stanford Coding Course Slashes 85% of Content in a Year, Final Project Now Only Half the Grade

A netizen went through the Stanford CS146S 'Modern Software Development' Fall 2026 and Fall 2025 syllabi week by week and gave a pretty brutal estimate: 85% of the content has been cut. 85% is not an official number; it's what this netizen said after comparing week by week. But looking at the changes, it really doesn't look like minor tweaks.

This course is taught by Mihail Eric at Stanford, and the course website has both the 2026 and 2025 syllabi posted. A netizen listed several key changes:

• Week 1: From teaching how to write prompts to hand-coding a coding agent like Claude Code in 200 lines of code.
• Final project: Weight dropped from 80% to 50%.
• Real open-source PRs: From 0 to 30%.
• Last year, a full week was devoted to AI IDEs, the Warp terminal, and generating apps with a single prompt; this year they're removed.
• New additions: Agent Skills, CLAUDE.md / AGENTS.md / hooks, agent-ready codebases, background parallel agents, MCP portal, model routing, and the last week covers software factories.

The course description repeatedly mentions MCP, agent skills, spec-driven development, loop engineering, and software factory. In plain terms: teaching how to feed context to agents, write product requirements as executable specs, and design processes where humans and agents repeatedly plan, build, and evaluate.

The grading change is the most noteworthy. A netizen pointed out that the final project was cut from 80% to 50%, with 30% for real open-source PRs, which is equivalent to changing from 'Can you build it?' to 'Are maintainers willing to merge your code?' When code is cheap, what's tested is the ability to read someone else's codebase and follow their rules. In a university course, this is fresher than spending a few more weeks on prompts.

If you only want to watch one class, Week 1 has the best value. The 9/22 session hand-writes a coding agent in 200 lines, with slides and complete code public. The core isn't complex: four tools—read, write, edit, bash—plus a loop. On 9/24, they dissect the system prompts of mainstream coding agents, and the original prompt files are also released.

The guest list also shows the course's focus. Lee Rob, former VP of Developer Relations at Cursor, now working on model behavior at SpaceX, gives the first talk. Boris Cherny, father of Claude Code, returns to share. Mihail asked him how many Claude agents he runs simultaneously, and he said 'thousands.' On 10/22 it's EnoReyes, CTO of Factory, followed by speakers from Cognition, Semgrep, Cloudflare, Replit, etc., including Amjad Masad and Elad Gil.

The 30% open-source component is not just nominal. The course website lists 15 open-source partners, including Vercel, Semgrep, OpenHands, Unsloth, etc. Students submit PRs to these projects, directly into production-level codebases. A netizen said this is a first for a university course.

The syllabus shrinking in a year is easy to see as less content. But most of what was deleted is the middle layer that was taught last year and has now been eaten by models and standards: how to click around an AI IDE, how to use the terminal, how to generate an app with one prompt. What remains is more fundamental—four tools plus a loop—and also more high-level: AGENTS.md, MCP, model routing, cost optimization. That middle layer of tools changes every six months; teaching it means it's outdated anyway.

For non-programmers, the four tools plus a loop in Week 1 are worth watching. A netizen said that after hand-coding an agent loop once, they finally understood what those IDE buttons do. The later weeks on MCP routing are more useful than another lecture on prompts.

Course materials are being released on themodernsoftware.dev gradually, so non-Stanford students can follow along. Even syllabi can't keep up with AI's pace; teachers are more anxious than students. This netizen comment fits well here. If you want to follow, just start from Week 1.