The AI Engineer YouTube channel just crossed one million subscribers. Thank you to everyone who has watched a talk, shared it with a colleague, or applied what they learned in production. For this special episode, we invited a guest to unpack a question many of us are asking at work: what should software engineers actually be learning right now?
Mihail Eric has two practical proving grounds for his answer. He teaches The Modern Software Developer at Stanford and leads AI at Monaco, where engineers use coding agents to ship software and debug production incidents. I asked him to walk us through the syllabus and share what has held up against real-world testing across his own engineering team.
Consider this a primer for anyone heading to AIE Code in San Francisco next week, serving as essential background reading for your upcoming conversations with speakers, sponsors, and attendees. If you are reading this as a Latent Space subscriber and haven’t grabbed a ticket yet, you can use this discount code for 30% off (sorry, no refunds).
Last Year’s Homework Can Now Be Solved in a Single Prompt
The first iteration of CS146S featured nine assignments and a capstone project. In one exercise, students had to run an AI code reviewer on a repository and evaluate its feedback. By the next iteration, Mihail found that students could one-shot most of those exercises. When he polled a room full of Stanford sophomores and juniors about their coding habits, every single student surveyed was already using agents daily.
So this time around, part of the coursework shifted to contributing to open-source projects, where repository maintainers review the students’ patches. I love what this shift accomplishes. You can indeed have an agent implement a feature in an afternoon, but you still have to figure out whether the maintainers actually want it, why the existing code was written the way it was, and what else your patch might break. That back-and-forth is also precisely how junior engineers learn from seniors.
There is an exercise from this episode that you can steal for your next PR (Pull Request). Geoffrey Litt’s explain-diff skill turns a code diff into a guided visual tour and then quizzes you on the changes. Try running it against a patch your agent just produced: which questions force you to reopen the code before you can answer? Watch the discussion.
Monaco’s Buy-vs-Build Strategy
At Monaco, the team evaluated several code-review products and chose to stick with Devin. But for agent workspaces, they took the opposite route. Background agents need an environment to run the product, reproduce bugs, and test patches. The specifics depend heavily on how Monaco’s cloud environment is configured, so the team built their workspaces in-house. Mihail’s rule of thumb: “You buy to learn, and you build to keep.”
One result of this work is Hal. An engineer can ask it to debug an issue and submit a PR; a salesperson can ask the exact same system how the product works under the hood. In the former case, the valuable artifact is a reviewable code change; in the latter, it’s an explanation ready for a customer. If you dump a screenful of debug logs on both of them, you’re only serving one of those audiences.
One question I keep coming back to is: what happens on the other end of the PR? An agent only sees the codebase as it exists today. But where does it look to find out why the team rejected the exact same change six months ago? This connects everything from Basis’s Authority Map to Mihail’s prediction that open-source repositories will evolve into agent-run software factories. Which is also a compelling reason to put students in front of maintainers today—someone still has to decide what truly belongs in the project.
If you’re building with coding agents, come say hello at AI Engineer CODE. And if you’re teaching this material yourself, I’d love to hear which assignment you ended up rewriting first.
Show Notes
- Stanford CS146S — The ten-week syllabus, assignments, and open-source partners. You can also toggle back to the previous academic year for comparison.
- Building Coding Agents and Demystifying Claude Code — Mihail’s deep dive into the tool loop and the harness surrounding it.
- Understanding is the New Bottleneck — Geoffrey Litt on making sense of AI-generated diffs.
- Monaco’s Agent-Developer Workspaces — The team’s thinking on building environments for agents to run and test applications.
- How Code Review Dies and the Faros study behind it — Heavy AI adopters merged 98% more PRs, but PR review time spiked by 91%.
- Basis: How we made our monorepo ergonomic for agents