Stanford’s CS146S syllabus teaches AI-native development through open-source practice

From scattered blogs to a formal curriculum

Mihail Eric launched the first iteration of CS146S in fall 2025 after months of watching coding agents move from prototype to daily workflow. Students must use an agent to implement a feature, then submit a pull request that a real maintainer reviews. This shift forces learners to confront the social and technical dimensions of code contribution rather than merely generating syntactically correct output.

Monaco’s “buy to learn, build to keep” in practice

At Monaco, the AI team evaluates external code-review products but ultimately builds its own agent workspaces. The rule of thumb, quoted by Eric, is “You buy to learn, and you build to keep.” The purchased product, Devin, serves as a learning platform, while the internally built workspaces host background agents that reproduce failures, run tests, and open PRs. One outcome is Hal, a dual-purpose assistant that can investigate a breakage for an engineer or generate a product explanation for a salesperson. The design principle—avoid giving both audiences a raw wall of debugging output—illustrates how internal tooling can be tailored to distinct stakeholder needs.

Open-source repositories become agent-run factories

A recurring theme in the interview is the question of provenance: when an agent proposes a change, how does it discover why a similar change was rejected six months earlier? Eric points to Basis’s Authority Map as a framework for agents to locate historical decisions, intent, and behavior within a repository. He predicts that open-source projects will evolve into “agent-run software factories,” where maintainers act as the final gatekeepers of change. By sending students to interact directly with maintainers, the course embeds this reality into the learning loop, ensuring that future engineers appreciate the human judgment layer that AI cannot replace.

Understanding the diff as the new bottleneck

Geoffrey Litt’s “explain-diff” skill is highlighted as a core assignment. After an agent writes a patch, students use Litt’s visual walkthrough to answer targeted questions about the change. The exercise reveals that the primary obstacle is not generating code but interpreting AI-produced diffs. Supporting data from a Faros study—cited in the show notes—shows teams with high AI adoption merged 98 % more PRs while their review time grew 91 % longer, underscoring that review latency, not code creation, is the limiting factor.

Our read: The CS146S redesign captures the most pressing friction in AI-augmented development: the need for engineers to validate, contextualize, and communicate machine-generated changes. By embedding open-source contribution, diff comprehension, and internal tooling strategies into a single, repeatable prompt, the course offers a scalable template for other institutions seeking to train the next generation of AI-native developers.