Mihail Eric is replacing easy-to-automate homework with contributions to real software in Stanford’s CS146S. Students must understand unfamiliar repositories and defend changes to their maintainers. On Latent Space, Eric connects that teaching experiment with the coding tools his team uses at Monaco.
Make understanding part of the assignment
The earlier course had nine exercises and a final project. By its second run, Eric found that many exercises could be finished with one prompt. He says every student he surveyed already used coding agents, models running multi-step tasks with tools.
Open-source contributions add constraints that disposable assignments miss. A maintainer can reject a change, question its design, or explain why an apparently reasonable implementation breaks another part of the system. Eric wants that exchange to provide the mentorship junior engineers risk losing when agents handle their initial tasks.
Computer-science fundamentals remain prerequisites. This is training in applying engineering judgment with agents, not a proposal to replace algorithms courses. The partnership model was only beginning when the interview was recorded. Its educational results are still ahead. 10:02–15:04
Put the design decision before the code
Eric’s workflow for large codebases starts with research. Ask the agent to inspect the relevant system and propose alternatives. A human chooses an approach before the agent writes a detailed plan and begins implementation.
That sequence makes the design decision visible while changing direction is still cheap. It also creates something concrete to review beyond a request to “build the feature.”
Context engineering means choosing which information an agent receives and when. Delegation makes it harder. A subagent may need the parent’s history, a focused brief, or a condensed account of earlier work. The speakers describe this as an unresolved engineering problem. Dumping every document into every conversation can make incidental remarks look like permanent requirements. 32:18–38:15
Skills, reusable instructions for repeated work, introduce a similar maintenance burden. Eric favors workflows that explain a code change and quiz its human reviewer. The hosts warn about bloated instructions and skills that activate in unrelated projects. Test when a skill should stay silent, too. 38:15–46:42
Separate Monaco’s installed tools from the syllabus
Eric names three parts of Monaco’s engineering setup:
- Monocoder runs background coding work in isolated environments. Monaco built it on Coder to support long-running and parallel tasks.
- Hal lives in Slack. It investigates failures, proposes fixes, and helps salespeople understand the product. Its explanations change with the audience.
- Devin supplies purchased code review. Eric says engineers adopted it because they found it useful, after earlier reviews of other products disappointed them.
His maxim is “You buy to learn and you build to keep.” It describes a willingness to learn from outside products before owning infrastructure that needs deeper customization. Devin remains an exception to building internally. The host also discloses an advisory relationship with its maker, Cognition. 59:21–1:05:24
Monaco’s customer-facing sales platform is a separate system. Eric says it has a bespoke agent harness, software that manages model execution and tools. It supports targeting, meeting preparation, and pipeline work. Parallel and Exa supply search alongside Monaco’s own enriched company-and-contact data. Claims about superior sales performance remain the vendor’s account. 50:23–59:21
Fix verification before multiplying agents
Eric treats repository readiness as a prerequisite for parallel agents. Clear boundaries, useful documentation, and executable checks give each agent a way to discover whether its change works.
Some companies have a simpler problem: slow builds, crowded merge queues, or applications that take too long to start. Another agent does little for those constraints. 1:05:24–1:12:37
Review creates a second limit. Monaco writes detailed pull-request descriptions, summaries accompanying proposed code changes. They explain implementation and point reviewers to affected files. The interviewer raises a harder question: who preserves the whole product’s intent across hundreds of individually plausible patches?
A versioned record of current behavior and earlier decisions might help. The discussion does not show that Monaco has solved this. Nor does it establish a general rule for safely approving changes automatically. 1:12:37–1:20:37
Keep ownership of risk and outcomes
Security needs explicit attention to dependencies and permissions. The speakers briefly entertain rebuilding dependencies internally, then challenge the assumption that private code is safer. Fewer outside packages can also mean fewer outside reviewers.
Authentication identifies a user. Authorization decides what that user may do. Engineers need that vocabulary to specify safe behavior, especially around customer data and billing. 1:20:37–1:30:13
The software-factory vision combines agents, checks, and feedback into a system that builds software. Eric places current human control around intent and review. Classroom plans involving factory infrastructure are experiments, not proof that unattended development is solved.
The closing advice is practical: deliver products, build evaluations of actual behavior, and publish rigorous engineering writeups. Elaborate orchestration that mainly produces more orchestration is a poor substitute. 1:41:22–2:00:43
Source: Latent Space’s full interview with Mihail Eric, published November 8, 2026. Watch on YouTube.