The newest edition of my Stanford course — 9 months in the making. 85% of the Fall 2025 material is gone, replaced by what AI-native engineers actually need: agent skills, context engineering, MCP, agentic review, software factories. All resources will be available online.
— @mihail_eric September 2, 2026
Engagement context: the September 2 announcement drew roughly 9,500 likes on X as of September 7 (our collection date) — unusually high for a syllabus announcement, and a signal of how much demand exists for AI-native engineering education.
What Changed for Fall 2026
Two structural changes stand out. First, the syllabus pivot: the Fall 2025 edition taught LLM foundations, context engineering, and tool workflows; Fall 2026 assumes that base and rebuilds everything around agents as the primary builders — with humans providing specs, review, and taste. Second, the partnership model: students ship weekly pull requests to production-grade open-source AI repositories, with partner support and mentorship. Eric's framing: this has never been done before in a university course.
Note precisely what is confirmed: the announcement text says “All resources will be available online.” Claims that classes are livestreamed circulate in secondary posts, but the confirmed fragment covers resources only — treat streaming as unverified until the course site states it.
The Fall 2026 Syllabus
Confirmed Fall 2026 topics (course site + announcement): agent skills, advanced context engineering, MCP portals, agent-ready codebase principles, agentic code review, security, parallelizing background agents, and software factories. The course description frames the outcome precisely: design effective agent-driven workflows, compose tools and skills into reliable development systems, and apply software-factory principles at speed and scale. A week-by-week schedule had not been published as of September 7 — the Fall 2025 archive and instructor posts are the best guide to structure until it lands.
| Topic area | What it covers |
|---|---|
| Agent skills | Composing skills into reliable agent systems — the course treats skills as first-class engineering artifacts. |
| Advanced context engineering | Providing agents the right context and capabilities across complex projects. |
| MCP portals | Model Context Protocol integration as course content, not an aside. |
| Agent-ready codebase principles | Structuring repositories so agents can contribute effectively. |
| Agentic code review | Review workflows where agents and humans share the loop. |
| Security | Permissions, gating, and the risks of delegating authority to agents. |
| Parallelizing background agents | Running multiple agents without losing coordination. |
| Software factories | Iterative workflows where humans and agents plan, build, evaluate, and improve together. |
The OSS-Partnership Model
A July 2026 call from the instructor named the first partners: Pi (the coding harness powering OpenClaw), Vercel, Warp, marimo, Unsloth AI, and Milvus (Zilliz), “and many others,” with 100+ students contributing weekly. The final Fall 2026 roster was not published as of this writing — the course homepage lists partners as logos, and the July post is the last named list. The mechanics per the announcement: students ship production-grade pull requests to partner repos, with support, training, and mentorship from the projects themselves.
How to Access the Materials
Three practical paths. Follow the course site — themodernsoftware.dev carries the syllabus, readings, and assignments. Use the assignments repo — github.com/mihail911/modern-software-dev-assignments (Fall 2025 assignments are up now; ~3.8k stars as of September 7, Python 3.12 + Poetry). Watch the Fall 2025 archive — the full 10-week breakdown, slides, and readings from the inaugural year are public. Fall 2025 classes were not recorded; expect the same for Fall 2026 until stated otherwise.
Fall 2025: The Inaugural Year
Historical context, from third-party guides corroborating the archive: the 10-week arc ran LLM foundations, coding-agent anatomy and MCP, context engineering and PRDs, agent patterns and Claude Code, the Warp terminal, testing and security with Semgrep, code review with Graphite, UI and build with Vercel, DevOps with Resolve AI, and the future role of the engineer. Guest speakers included Boris Cherny (Claude Code creator), Cognition's Silas Alberti, Warp's Zach Lloyd, Semgrep's Isaac Evans, Vercel's Lee Garcia, and a16z's Martin Casado — the Fall 2026 roster is unpublished, so treat that list as last year's, not this year's.
Who Should Follow Along
Working engineers upgrading to agent-native workflows — the syllabus maps almost one-to-one onto the skills gap employers describe. Team leads designing AI-native engineering orgs — the software-factory and review-loop content is org-design material. Students — the OSS-partnership model means real production contributions, which is the strongest portfolio signal available. For Stanford students: 3 units, CS111/CS161-equivalent prerequisites, an application deadline of 9/27 per the bulletin, and classes on campus from 9/22.
A Self-Study Path From the Public Materials
If you are not enrolled, this is the usable sequence from what is public today (Fall 2026 week-by-week lands when the course publishes it):
- Weeks 1–2: foundations. Work through the Fall 2025 archive's LLM-foundations and agent-anatomy readings; install and use at least two coding agents (one proprietary, one open — e.g. Claude Code and OpenCode) on the same repo.
- Weeks 3–4: context engineering. Study the archive's context/PRD materials, then write a spec for a real task and hand the same spec to both agents — compare drift.
- Weeks 5–6: MCP + skills. Build one MCP server and one agent skill; wire both into an agent and measure the difference in tool-call quality (the Fall 2025 Semgrep/security materials cover the safety side).
- Weeks 7–8: review + security. Run agentic code review on your own pull requests; study the archive's Graphite/Semgrep weeks for the review-loop patterns the Fall 2026 syllabus rebuilds.
- Weeks 9–10: the real thing. Ship an actual pull request to one of the named partner projects (Pi, Vercel, Warp, marimo, Unsloth, Milvus) — the course's core mechanic, available to anyone.
Source discipline: steps 1–4 map to corroborated Fall 2025 materials (third-party guides + archive); step 5 maps to the confirmed Fall 2026 partnership model. Treat the weekly mapping as reconstruction, not the official syllabus.
FAQ
Can non-Stanford students take CS146S?
Not for credit — but the materials are public. The confirmed statement is that all resources will be available online; the Fall 2025 archive and assignments repo are already usable self-study paths.
Is there a professional version?
Yes — the instructor runs a Maven course (“AI Software Development: From First Prompt to Production Code”) and a leaders' course on AI-native engineering teams. Those are separate paid offerings; the Stanford class is the source of the public materials.
What if I can't contribute to an OSS repo?
The weekly-PR model is Stanford-enrollment mechanics. For self-study, the equivalent is shipping real pull requests to any of the named partner projects (Pi, Vercel, Warp, marimo, Unsloth, Milvus) — the assignments repo provides the structured practice path.
Sources
- themodernsoftware.dev — CS146S course site (Fall 2026 topics, format, partners)
- Stanford Bulletin — CS146S (units, prerequisites, application deadline)
- GitHub — modern-software-dev-assignments (public assignments)
- Mihail Eric — instructor page (course + professional offerings)
- themodernsoftware.dev/fall2025 — inaugural-year archive
Get the latest on AI, LLMs & developer tools
New MCP servers, model updates, and guides like this one — delivered weekly.