How Stanford Teaches AI Coding Now

Mihail Eric says week five of Stanford’s ten-week “Modern Software Developer” course is underway, alongside a newsletter with the same theme. It is his second year teaching the course, but he discarded roughly 80% of last year’s material after model capabilities changed dramatically around November–December 2025, especially with Claude Opus 4.5 and the rapid adoption of Claude Code over the holiday break. The revised curriculum now covers agent harnesses, loops, software factories, automated code review and approvals, merge conflicts, DevOps, and preparing codebases for agents.

Instead of assigning invented exercises, Mihail arranged for students to contribute to major open-source repositories as homework. He jokes that maintainers may “love slop,” but says a student receives a zero if a maintainer complains about low-quality generated code. His students have strong computer-science fundamentals, but the course teaches them to onboard quickly to large repositories, understand architecture, scope issues, gather context, choose MCP tools, and use skills appropriately. The core rule is simple: students must read every line their agent produces rather than blindly merging its output.

Mihail pushes back on the idea that capable coding agents eliminate structured development. He teaches a version of the RPI loop—research, plan, implement, and test—which includes gathering context before coding, reviewing a reasonably detailed specification, and defining how outputs will be verified. The host connects this to Jesse’s popular Superpowers project, whose continued popularity after a year of rapid model advances suggests that disciplined agent workflows remain valuable. Mihail says the best use of his own time is setting up verifiers that can independently check what an agent has done.

The host describes Jesse’s “prove it” skill, which makes an agent record a Loom-style walkthrough showing that the application works. Mihail says developers are also experimenting with more informative code diffs, including visual architecture diagrams inspired by older technical formats, plus screenshots and recorded flows. Playwright with MCP can provide browser-based checks, and a recently merged Web MCP debugging capability lets an agent operate on a web application specifically to diagnose problems. Mihail expects verification companies to expand into broader software factories because closing the loop requires understanding both upstream development and downstream validation.

Mihail’s Stanford guests included Boris, the creator of Claude Code, and Lee Robinson, formerly a Cursor developer-relations executive and now at SpaceX. Boris described running perhaps thousands of agents and argued that a one-off prompt is a mistake: the agent should talk to a loop and prompt itself through meta-prompting. With no token-cost constraint, he experiments by intelligently brute-forcing difficult problems. Lee delivered a dense, 45-minute explanation of modern agent harnesses, covering computer use, context engineering, tool calls, subagents that simulate infinite context, and compaction.

The conversation compares products such as Grok bots, Instinct, and Dots. The host finds Instinct substantially more proactive—for example, asking how a movie was—while Dots mainly surfaces potentially useful information from email. Lee explained that agent systems use file systems, regularly compact conversation state, and write information out so it can be retrieved at the right moment. The speakers also discuss new approaches to agent sandboxes and file systems, along with efforts to replace or rethink Git. Anthropic has even published work on rebuilding version control because conventional platforms struggle with the volume and reliability demands of agent-generated commits and pull requests.

Mihail says he hears one or two startup pitches a week and stops paying attention when the same idea keeps returning. A common pitch adds JavaScript to an application, traces model requests and responses, and routes each task to a cheaper specialized or open-weight model. He agrees that companies want to keep increasing usage without allowing token costs to rise, but says routing and prompt or skill optimization can address costs without another layer of software. He calls company traces “the new oil” because they capture customer and application interactions, while warning that the category is difficult to differentiate.

At Monaco, Mihail leads development of agents for go-to-market work, including pipeline reviews, total-addressable-market building and segmentation, demand-generation recommendations, and outbound campaigns. Monaco combines CRM functions, data enrichment, and outbound workflows, then uses emails, meeting recordings, transcripts, Slack messages, uploaded collateral, and public announcements to infer an organization’s tone, ideal customer profile, target buyers, pain points, and relevant timing. Its “Monaco memory” is a persistent store of what the system knows about an organization and its users. Rather than being purely self-serve, every customer gets a forward-deployed account executive to help configure the system.

Monaco has built its own go-to-market agent harness, while outsourcing undifferentiated infrastructure such as micro-VMs to Vercel sandboxes. Natural-language instructions are converted into an intermediate representation, effectively JSON or YAML describing a workflow, then compiled into a graph that can loop and self-heal. Mihail says the defensible pieces include the data and context layers, memory, compaction, tools, and Monaco’s position as both system of record and system of intelligence. MCP tools should be designed around agent intent rather than one-to-one API endpoints: a meeting-search tool, for example, should support both keyword and semantic queries through one meaningful entry point.