AI raises entry-level engineering standards to near-senior level
AI tools now handle foundational coding well enough that companies need architects who can weigh design trade-offs; new graduates' baseline skill bar has risen by a tier.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Principles stabilize: teach methodology, not case studies
Year one courses broadly survey how AI applies across the software development lifecycle; year two narrows to principle codification—systematic thinking about context engineering, design, and architecture. This shift mirrors AI engineering maturing from wild growth into accumulated best practice. Course syllabi need annual rewriting since last year's hot company may vanish, but underlying principles are now stable enough to teach as methodology rather than case studies.
— Mihail EricAI erases entry-level positions; newcomers must now architect, not just code
Graduate hiring expectations have risen from entry-level to near-senior competency—not market inflation, but real change driven by AI. Coding agents handle foundational work so well that the place where newcomers most often stumble is now automated. The floor disappeared, not because the ceiling rose. Companies now need architects who can weigh design trade-offs. When basic coding is off the table, entry-level positions as a category cease to exist.
— Mihail EricReal projects restore the mentorship chain through production code review
Year two's core innovation: students work directly on real open-source projects—million-line production codebases, not simulated exercises—collaborating with maintainers. This solves a critical problem: mentorship. When AI takes over foundational coding, the knowledge transmission between junior and senior engineers that once flowed through "the process of coding" loses its vehicle. Real code review re-anchors that inheritance chain.
— Mihail EricRPI framework preserves human decisions while scaling to production codebases
The RPI framework (Research-Proposal-Implementation) isn't concept theater—it's a practical tool for million-line repositories. The logic: Research phase maps the codebase context; Proposal phase surfaces multiple approaches for humans to decide between; Implementation phase executes. This design keeps humans in the decision loop while the tool handles genuinely complex real tasks.
— Mihail EricBuy first to learn, build only when you need control
Monaco's tool philosophy: buy to learn, build to keep. Early on, they tried multiple CodeReview platforms to validate what they needed, then self-built to retain control. They built Monocoder (based on coder.com) and Slack bot Hal because existing SRE tools lacked maturity. CodeReview was the exception—they chose Devin and stayed with it, because it crossed the usability threshold.
— Mihail EricCode generation now faster than human review; triage beats line-by-line audit
AI can theoretically produce 30+ PRs a day, but human review throughput hasn't scaled proportionally. If review stays traditional line-by-line inspection, reviewers become the hard bottleneck. Mihail sees the middle ground: triage by risk. Some PRs auto-approve; some need human eyes; and even when humans review, it might be spec-level, not character-by-character.
— Mihail EricRisk-based review: scrutinize auth and payments, trust everything else
Auth, Billing, Data Storage—three zones marked "no casual code", requiring careful design and review. Everywhere else: "generate freely". This tiering matters because AI-generated code's main risk isn't more bugs—usually fewer—but "you don't know what you wrote". When review capacity can't cover everything, deep inspection on high-risk zones and deference elsewhere becomes the pragmatic trade-off.
— Mihail EricManager of Agents replaces traditional engineer role, orchestrates system health
Engineering metrics are shifting. Today it's PR merge rates; in two years, Eric thinks that standard will look absurd. He sketches a new role: "Manager of Agents"—not people management, but orchestrating an agent system, requiring understanding of product goals and technical constraints, decision-making across abstraction layers, setting goals, course-correcting. Even inside Anthropic, the product head and engineering head roles haven't merged yet.
— Mihail EricIn their own words · checked verbatim
It's like you have to generally be better than, you know, almost at the senior level going into this.
Mihail Eric6:27
In the course of being in all those projects, you'll actually be talking to the maintainers, they will be reviewing your code, they'll be helping you build mental models of, like, good software.
Mihail Eric13:00
there is a way to do this that actually does scale out to code bases that are, you know, millions of lines long.
Mihail Eric34:14
you buy to learn and you build to keep
Mihail Eric1:01:53
you could in theory like put up 30 plus PRs a day, uh, and then what?
Mihail Eric1:13:40
The parts of the code base where security is very important is data storage
Mihail Eric1:28:49
your entire job is making God boxes and then taking them apart
Mihail Eric1:36:26
you write prompts that write prompts
Mihail Eric1:47:55
Figures
| Claude Code system prompt reduction ratio | 80% | 29:59 |
| Basic agent loop code lines | 200 lines | 26:45 |
| Claude Code original system prompt size | 20k token | 29:59 |
| Total course weeks | 10 weeks | 20:19 |
| Monaco funding stage | Series B | 51:12 |
| Monaco primary customer range | Seed to Series A | 52:16 |
| Theoretical daily PR generation capacity | 30+ | 1:13:40 |
| Anthropic's stage in cloud adoption model | Stage 3 (assisted) | 1:57:30 |
| Boris cloud adoption model levels | 5 levels (gated, assisted, parallel, supervised autonomy, AI native, counting from 0) | 1:56:26 |
Glossary
- RPI framework
- Research-Proposal-Implementation; a three-step context engineering method for handling million-line codebases.
- sub-agent
- An independent agent spawned from a parent agent to handle subtasks, with inherited or isolated context.
- God box
- A central middleware component through which multiple agents route all requests and data, a critical infrastructure design challenge.
- context engineering
- Systematically organizing and supplying context information to agent systems, rather than writing scattered prompts.
- software factory
- An open-source project operated primarily through continuous agent-driven code generation and updates.
- Manager of Agents
- A new engineering role focused on orchestrating multiple agent systems, not managing people, requiring understanding of both product needs and technical constraints.
How to listen
Engineers focused on AI-native engineering practices, tech managers, and founders wanting to understand how AI startups decide whether to buy or build internal tools.
The opening section on specific course selection and curriculum week-by-week details has lower information density and can be skipped.