[One Million AI Engineer Special] How Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco

Stanford's course treats AI engineering as a discipline

Mihail Eric teaches Stanford's CS146S, a course built on the premise that coding agents are easy to build and hard to use well. The scarce skills are context engineering, tool design and keeping a code base legible enough for agents to verify their own work. As writing code gets cheap, the constraints move to reviewing it and to saying precisely what you want.

Takeaways

• A coding agent is about 200 lines. Week one has students build one from scratch using only introductory algorithms knowledge, then uses a MITM proxy to intercept Claude Code's requests and show the real system and compaction prompts.
• Context engineering and MCP stay central. Mihail Eric is pro-MCP and covers context engineering at scale and tool design. At scale an MCP portal becomes a single point of entry for governance, security, auditing and cost optimisation.
• Agent-ready code bases set the ceiling. Week five covers documentation and service-level boundaries, and Factory shipped readiness tooling early. How verifiable and understandable a code base is to an agent decides what multi-agent systems can do.
• Code review is the new bottleneck. Week six covers agentic code review, drawing on posts by Cloudflare and Amplitude. Mihail Eric says a fully automated system has two choke points, code review and intent at the top.
• Software factories are only 30 percent agent-driven. Even teams building software factories say around 30 percent of contributions are fully agent-driven. Selling a whole factory means landing as a coding agent first, then expanding into background agents and SRE.
• Monaco buys to learn, builds to keep. The Series B company built its own agentic harness for sales workflows such as TAM and ICP scoring, and kept Devin for review after surveying and dropping tools such as CodeRabbit.

How the course is run

• Mihail Eric had the idea in November 2024, when knowledge about coding agents was scattered across blog posts, and it took nine to ten months to clear the university committee.
• The first version had nine assignments and an open-ended capstone, which he found too easy because agents could one-shot them.
• Students include Stanford postdocs from pathology and medicine plus MBA students, and final projects use Y Combinator's request for startups as seed ideas.
• Open-source partners including Warp and Vercel give students real code bases to work in, and the class is not for beginners or a replacement for traditional computer science.

Career advice and hype cycles