The short version
- Computer science graduates face a 6.1% unemployment rate. Companies are hiring seniors who need no ramp-up and letting AI cover the rest.
- The roles shrinking fastest all share one trait: someone else already wrote the spec. Ticket-following engineers, manual QA, routine analysts.
- The three roles growing fastest — Forward Deployed, AI Platform and AI Feature Engineers — are the ones that write the spec, build the foundation, or turn a messy human need into software.
- The four skills that put an engineer in the top 1% — synthesis across disciplines, task decomposition, directing AI agents, and building things that can be tested — can all be started long before college.
- Our Robotics & AI Mentoring program, grades 4–12, is built around the early version of every one of them.
In a recent interview, Mihail Eric — AI lecturer at Stanford and a former head of AI — gave one of the clearest pictures yet of what AI is actually doing to software engineering jobs. It isn’t the picture most families have. Software engineering isn’t disappearing. It is splitting in two, and the split runs along a line a student can start preparing for in middle school.
The market no one warned students about
The headline number is uncomfortable: computer science graduates now face an unemployment rate of about 6.1%. Eric’s explanation is not that CS became useless. It is that companies are making a short-term trade. A senior engineer needs no onboarding; a junior one does. With AI tools covering a lot of routine work, many employers have stopped paying for the ramp-up.
Eric thinks that’s a mistake — juniors have what he calls “high slope,” they learn faster than anyone on the team — but it is the market students are graduating into. His conclusion is blunt: a CS degree on its own is no longer a guaranteed path to a job. What separates the graduates who get hired are meta-skills: scoping a project, communicating, and solving problems that cross disciplines.
Which roles shrink, and which explode
Eric’s list of threatened roles looks at first like a list of junior jobs. Look again and it’s a list of spec-following jobs:
- the software engineer whose work is “hands-on-keyboard ticket cranking” from a Jira description,
- the QA engineer whose job is clicking through buttons and triaging,
- the entry-level consultant assigned well-defined manual tasks on an existing project,
- the analyst running routine data exploration with no question of their own behind it.
In every case, a person upstream has already decided what to do. AI coding agents are now very good at the how.
Where the jobs are moving
The spec-followers shrink. The spec-writers grow.
Roles Mihail Eric expects AI coding agents to displace, next to the three he expects to explode.
Shrinking
-
Ticket-following software engineer
Turns a Jira description into code — the spec is already written.
-
Manual QA engineer
Clicks through UI buttons and triages against a checklist.
-
Entry-level outsourced consultant
Carries out well-defined manual tasks on someone else’s project.
-
Entry-level data analyst
Runs routine exploration with no question of their own behind it.
Growing
-
Forward Deployed Engineer
Sits with the customer and turns messy, unstated needs into a spec.
-
AI Platform Engineer
Builds the foundation agents run on — hosting, inference, evaluation, architecture.
-
AI Feature Engineer
Builds domain-specific agent features on top of that platform.
The three roles he expects to explode are the mirror image.
Forward Deployed Engineers sit on-site with customers and act as the bridge between them and the technical team. The core skill isn’t writing code — it’s hearing messy, contradictory feedback and noticing the need the customer never actually said out loud, then turning it into something buildable.
AI Platform Engineers build the foundation AI agents run on: model hosting, inference providers, evaluation frameworks, system architecture. This is deep systems work — the kind that requires understanding how computers, networks and software actually fit together.
AI Feature Engineers build specialised, domain-specific agent features on top of that platform — Eric’s example is an insurance-focused document summariser. The skill is knowing a domain well enough to know what “good” looks like in it.
The four skills that put an engineer in the top 1%
Eric names four skill sets that separate the top 1% of engineers from everyone else. None of them is “write code faster.”
- Interdisciplinary synthesis — combining programming with a non-engineering domain like design, hardware or product, and building end to end.
- Task decomposition — breaking a complex problem into structured subtasks and dependencies an AI agent can execute.
- Agent orchestration — managing a fleet of AI agents like a scrum master managing eager interns: directing them, switching between them, catching them when they’re wrong.
- Agent-friendly infrastructure — codebases with real test coverage and clean builds, so agents can commit code without silently breaking things.
Here is the part that matters for a parent: every one of these has a version a twelve-year-old can practise. Not the job — the habit underneath it.
The top 1%, at twelve years old
Four skills, and what each one looks like before college
Mihail Eric’s four skills for standing out as an engineer, translated to middle and high school.
Interdisciplinary synthesis
Combine code with design, hardware or product to build the whole thing end to end.
At 12–17
A robot that works needs mechanical, electrical and software decisions to agree with each other.
In the program
Six engineering disciplines in one project — wiring, mechanics, microcontrollers, code, models and the demo.
Task decomposition
Break a hard problem into subtasks and a dependency tree an agent can execute.
At 12–17
“Make the car avoid obstacles” becomes: read the sensor, filter the noise, decide, steer — in that order.
In the program
Every project is scoped by the student with a mentor, and split into what the AI can help with and what needs their own judgment.
Agent orchestration
Direct a fleet of AI agents like eager interns, and catch them when they are wrong.
At 12–17
Using AI to draft code or explain a concept, then checking whether the answer is actually right.
In the program
AI is a tool in the student’s hands; the mentor’s job is to make sure the student is still the one thinking.
Agent-friendly infrastructure
Codebases with real tests and clean builds, so an agent can commit without breaking things.
At 12–17
Knowing how you will prove it works before you build it — and testing on real hardware, not hoping.
In the program
Hardware is an unforgiving test suite. The robot either scores or it doesn’t, live, in front of judges.
An honest word about competitive programming
If you have been following this debate, there’s a fair question for any program that trains students for USACO, ACSL and AP Computer Science: isn’t solving well-defined problems quickly exactly the kind of work AI is taking over?
For the job, partly yes. For the student, no — and the distinction matters. Competitive programming is how a student builds the algorithmic intuition to know why an approach will or won’t work, how long it will take, and where it will fail. That’s precisely the judgment you need to break a problem down for an agent, and to catch the agent when its confident answer is wrong. You can’t supervise work you couldn’t have done yourself.
So we treat it the way a musician treats scales: not the performance, but the reason the performance is possible. It lives in one session a week. The other session is the open-ended, hands-on project where the student decides what to build.
How our Robotics & AI Mentoring program builds this
4–12
grades
8
students per cohort, max
2×
sessions weekly, year-round
6
engineering disciplines
Robotics is interdisciplinary by nature. A robot that works is one where the mechanical, electrical and software decisions agree with each other. Our students work across all of them — circuits and microcontrollers, motors and gear ratios, code, sensors and ML models — on the same project. That’s Eric’s first skill, practised every week.
The student writes the spec. There’s no kit and no assigned game. Each student chooses their own project, scopes it with a mentor, and breaks it into pieces — which parts an AI tool can help with, which parts need their own judgment. That is task decomposition, and it’s the Forward Deployed Engineer’s core habit in miniature: turning a vague “I want it to do this” into something that can actually be built.
AI is a tool in the student’s hands, not a replacement for their thinking. Students use AI the way working engineers now do. The mentor’s job is to make sure the student can tell a correct answer from a confident-sounding one — the exact supervision skill Eric describes. (We wrote more about this in why mentorship, not more content, is the right answer in the AI age.)
Hardware is an unforgiving test suite. Software can look finished and still be wrong. A robot can’t. It scores or it doesn’t, live, in front of judges who ask questions the student didn’t rehearse. Students learn early to decide how they’ll prove something works before they build it — the instinct behind every good test suite.
The AI is built, not just used. Computer vision, ML models and agents, trained on the student’s own noisy sensor data. A twelve-year-old in our program built an object-detection robot car — onboard camera, vision model, real-time driving decisions. That’s an AI feature, built on a platform the student wired themselves.
The mechanism
One loop, run continuously from grade 4 to grade 12
Every trait in the section above is a by-product of this cycle. Nothing here can be lectured into a child — it has to be run.
- 1 Choose builds Agency
The student picks the project. No rulebook, no assigned game.
- 2 Build builds Craft
Real hardware and real code — sensors, circuits, models they wrote.
- 3 Break builds Resilience
It fails. Usually in front of someone. A mentor sits with them in it.
- 4 Ship builds Obsession
They fix it because they want the thing to exist, not because it is graded.
- 5 Defend builds Fluency
They demo it live to judges and answer hard questions unscripted.
Then back to step one, with a harder problem the student chose themselves.
For older students, the High School Tech Hero track goes wider across the systems these jobs sit on — networking, cybersecurity, distributed systems, cloud, AI & ML and physical AI — with an AI-first workflow throughout.
See the program in full or what our students have built and defended. Related reading: the four bets Stanford’s AI-era operators are making, and the majors they point to.
Aasquare Academy — Robotics, AI & Computer Science Mentorship. Grades 4–12. Sunnyvale & Cupertino, CA.
Remarks and figures attributed to Mihail Eric are drawn from a public interview on how AI is changing software engineering roles. Titles are as given at the time of the interview.