From Writing Code to Managing Agents. Most Engineers Aren't Ready | Stanford University, Mihail Eric
These notes were generated by AI and may contain inaccuracies.
There is emergence of a new class of engineer: the AI native engineer, where AI is the new language. This generation of junior developers entering the workforce will be the first of this shift. A single developer becomes a manager of agents. Adding more agents doesn't always create better systems; it can make worse systems if just let go to do whatever. Properly handling multiple agents is the last boss in a game; doing it well puts you in the top 0.1% of users today.
I'm Mihail Eric. I lead AI at an early-stage startup in San Francisco and teach a class at Stanford called The Modern Software Developer. It's the first class focused on AI across the SDLC. Within hours of announcement, it filled up with over 100 students trying to get in.
Huge momentum in software development with AI making its way into every part. Anecdote: Berkeley graduate applied to a thousand places, heard back from only two—not even interviews.
Three factors created a perfect storm in the software ecosystem:
- Around 2021, huge surge in hiring post-COVID; companies overhired, then massive layoffs reducing workforce by 20-30%.
- CS major graduates doubled or tripled nationally/internationally in last 10-15 years.
- AI became popular; employers consider hiring fewer AI native people to cover quotas.
This generation of junior engineers must have good fundamentals and be fully AI native.
AI native engineer has strong foundation in traditional programming, system design, algorithmic thinking, and competence in agentic workflows.
Build incrementally: Don't start with 10 agents at once like Boris from Claude; wrong outcome. Master one-agent workflow to build complex software, then add isolated tasks (e.g., second agent fixes logo, third updates header copy). Iteratively add agents when confident, understanding task boundaries.
Key skill: context switching between agents—like managing eager interns. Monitor them in terminal/IDE; switch to push stuck tasks forward while remembering context. This mirrors skills of good human managers; best at agents have managed human developers.
Agent-friendly codebase: Would an agent understand and contribute without breaking things?
- Agents test against tests as contracts for correctness; need sufficient test coverage.
- Keep READMEs consistent with code; outdated READMEs confuse agents.
- Avoid spaghetti code from iterative agent builds; agents compound errors quickly (misunderstanding in step 1 magnifies in step 2).
- Ensure initial codebase is robust: airtight design, testing, build, linting, style checking.
- Use consistent design patterns (e.g., same API for creating objects everywhere); inconsistency confuses agents like it would humans.
Difference defined by taste in good software; some have it innately, others develop it.
In class, requirements like building five flows get baseline; bonus/extra credit separates top performers who invest beyond 100% for robustness, expansion.
Top students built startups from projects, continuing post-class because they see more potential.
Experimentation key to AI native development. Example: Anthropic rewrites Claude every week or two using Claude itself, constantly iterating based on user feedback.
Build experimentation into workflows: Try tools yourself to see what works.
Senior developers resist AI tools, ingrained in 20-year habits. Juniors are sponges: naive, unscarred by industry hardships (e.g., healthcare), brave to tackle problems—perfect for startups.
Juniors adopt AI native skills fastest, nimble enough to succeed where seniors cannot.
CS teaches thinking for complex systems via algorithms—like math. Developers break/fix/expand/iterate; willing to customize, debug internals. Developer arrogance: Sees any problem, thinks software solves it; confident to fix with known tools.
Prompt Claude/Codeex to build, add features; month later, beautiful overengineered software launches to no users.
Rem Koning, professor at Harvard Business School, studies entrepreneurship and AI. World needs ability to allocate intelligence.
AI native embeds AI in product to work directly with customers, removing human from loop—key to AI native organizations.
Open question: What happens when AIs talk/collaborate? Needs for each other? Provocative; could birth trillion-dollar companies.
Keep EO in your library
Save the videos and channels worth coming back to, and find them again in one place.
