Latent Space

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

Published 2026-11-08 · Duration 2:00:45

Summary

This session, featuring Mihail Eric (Stanford/Monaco), details the shift toward 'AI-Native Engineering' and the evolution of the Software Development Life Cycle (SDLC) driven by autonomous coding agents. The curriculum emphasizes moving beyond simple prompting to advanced techniques like context engineering, multi-agent management, and building robust, agent-ready codebases. Key architectural challenges discussed include managing the new bottlenecks in code review, securing complex dependencies, and scaling background agent tasks within CI/CD pipelines.

Download summary

Key takeaways

  1. The Shift to Deep Principles in AI Engineering 2:34

    The focus of modern development education is shifting from a broad survey of AI's impact on the SDLC to mastering deep principles, such as careful context engineering and thinking through design and architecture for complex tasks. The goal is to train engineers to be highly productive members of the workforce, even when the code is largely generated by agents.

  2. Open-Source Contributions and Mentorship 5:26

    To combat the potential gap between senior engineers and juniors in the age of agents, the course emphasizes having students contribute directly to major open-source, AI-focused repositories. This allows students to interact with professional maintainers and build mental models of good software design at scale.

  3. New Bottlenecks in the SDLC 20:03

    As coding agents increase velocity, the bottlenecks are shifting. Key areas requiring human attention include code review (1:20:37), security and dependency tracking (1:20:37), and managing the complexity of background agents and event-driven tasks (1:05:24).

  4. The 'Buy to Learn, Build to Keep' Philosophy 16:54

    Monaco's approach highlights a strategic principle: 'You buy to learn, and you build to keep.' While adopting best-in-class external tools (like Devon for code review), the company builds proprietary internal platforms (like Monicoder) to maintain control and adapt to specific domain needs.

More digests on AI Engineering

Technical details

  • Agent Architecture and Harnesses 420s

    The modern coding agent is viewed as a 'harness'—a structured system (often requiring under 200 lines of code) that orchestrates LLM calls. This harness includes multiple components like system prompts, plan prompts, and compaction prompts, allowing for sophisticated, multi-step reasoning beyond vanilla prompting.

  • Context Engineering 680s

    Effective context management for large codebases requires structured processes like Research, Proposal, Spec, and Implementation (RPI). This is crucial for ensuring agents can ingest and understand millions of lines of code, moving beyond simple, single-turn prompts.

  • Multi-Agent Management 760s

    When fanning out work to multiple sub-agents, the challenge is maintaining context consistency. This requires careful management of context inheritance and consolidation (compaction) to prevent agents from making mistakes based on incomplete or conflicting information.

  • Background Agents and CI/CD 1220s

    Agents are increasingly used for event-driven tasks (e.g., running on triggers in a codebase or during CI/CD). This involves building systems that monitor logs and analytics for live production, moving beyond traditional IDE-based interactions.

Mentioned resources

Channel & topics

Watch on YouTube · Back to latest

This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.