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

Stanford's CS146S entered its second run with a revised structure, led by Mihail Eric, who teaches The Modern Software Developer at Stanford and leads AI at Monaco. The course is built around a problem that has become hard to ignore: when a coding agent can complete last year's homework in a single prompt, the assignment itself has to change. In this iteration Eric brought open-source maintainers into the course. Students must learn an unfamiliar repository, propose a change, and then explain the patch to the person who might actually merge it. The context was a Latent Space interview marking one million subscribers on the AI Engineer YouTube channel.
The ten-week syllabus is published in full in the show notes. The interviews and screenshares walk through which lessons survived contact with Eric's own engineering team at Monaco, and the timestamps map out the argument: what junior engineers need to learn now, open-source maintainers reviewing student code, a tour of the syllabus, and then the machinery underneath coding agents. The engineering segments cover the loop, tools and harness inside an agent, context engineering for a large codebase, and the skills that let an engineer actually understand a diff.
The maintainer-review design is the most interesting part for anyone who builds curricula or runs onboarding. This suggests the aim is to make the artifact accountable to a real reviewer rather than a grader. Learning an unfamiliar repo, forming a change, and defending it are the steps most likely to expose whether a student can reason about a codebase instead of pattern-matching. In practice it also pushes students toward the parts of the work that agents struggle to own: knowing why a change is correct, judging blast radius, and anticipating what a maintainer will object to.
Monaco's own setup is the second thread. The company built its agent workspaces rather than buying them, purchased its code review tools, and gave an on-call agent access to Datadog and GitHub. Eric's framing for those choices, as quoted in the interview, is that you buy to learn and build to keep. That distinction is a useful lens for tooling decisions. Buying vends a known capability quickly; building creates something tied to how the team actually operates and worth maintaining. In this reading, code review and observability are commodity picks, while the environment an agent runs in is treated as differentiated.
The timing appears designed to close the loop between diagnosis and response. An on-call agent with access to Datadog and GitHub can read telemetry and the changes that preceded it in the same place, which is the difference between a notification and a starting point for investigation. The rest of the discussion touches the pressure this creates: making repositories agent-ready, what happens when code review becomes the bottleneck, security, dependencies and permissions, background agents and Stripe's Minions, MCP portals and infrastructure convergence, building the software factory, what work is left for developers, and the shift from personal adoption to team autonomy.
Referenced material expands on the mechanics. One piece reconstructs a coding agent from scratch, another traces a real agent harness, and an essay on killing code review argues that generation can outpace review. A piece on an Authority Map describes how agents separate current behavior from earlier design decisions, which is the sort of distinction a maintainer makes instinctively when deciding whether a patch fits.
Why it matters: teams hiring juniors and teams rolling out agents face the same question the course answers, which is what human work remains once generation is cheap. In my reading, the transferable lesson is process design: requiring a proposal and an explanation from the person who would merge the change forces understanding of the diff and makes output auditable, while the Monaco split offers a rough heuristic for build-versus-buy decisions around agent infrastructure.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
What do you assign when a coding agent can finish last year’s homework in one prompt? In the second run of Stanford’s CS146S, Mihail Eric brought open-source maintainers into the course. Students have to learn an unfamiliar repo, propose a change, and ...