Senior Engineers Are the Next DRAM Shortage

There’s a deficit coming that nobody’s put on a glide yet: elder engineers. Every business on earth is concerning to desire one who can run a fleet of agents alongside judgement — and nearly nobody is making new ones anymore, since AI ate the entry-level activity juniors used to study on. You cannot conjure a elder in a quarter. We fair watched this exact movie alongside memory: DRAM prices went vertical since petition exploded and you can’t build a fab overnight. Senior engineers are next, and the lone hedge is to build your own, starting now. The finest item you can build in the agentic era isn’t your merchandise — it’s another Engineer. I have juniors on my team. This is my problem, correct now. Here’s how I’m thinking concerning it.

Stop hiring and mentoring younger engineers and you’ve quietly unplugged the device that makes elder ones. A few years out, you run dry — correct whenever everyone alternatively is dry too, and the cost of the few remaining goes through the roof.

I’ve written concerning the org diagram flattening from the middle, and concerning how the bosses are coding again. This article is the part those two were circling around: what happens to the bottom of the chart, and whose job it is to fix it. Spoiler — it’s yours, elder engineer. It’s mine. And most of us are concerning to get it incorrect in a extremely specific, extremely tempting way.

Yes, the Naysayers Are Right (Sort Of)

The destiny obtain goes akin this: AI eats entry-level engineering work, entry-level jobs dry up, the ladder loses its base rung, and a few years from now there’s a gap in the workforce shaped exactly akin the seniors we forgot to grow.

And… they’re right. That nonaccomplishment manner is entirely real. If you obtain the activity juniors learned on and path 100% of it to agents, and you do nothing else, that is precisely what happens. And this isn’t fair me predicting it. Martin Fowler and Thoughtworks pulled a area complete of CTOs, architects, and elder practitioners together in Engelberg this former June, and among the five header findings from that retreat was — I’m quoting the study — “an apprenticeship and skills-transmission crisis.” Not a maybe. A crisis, named as such, that came up independently in at smallest six distinct sessions. Their words for the nonaccomplishment mode:

If juniors never get to battle alongside genuine code, genuine manufacturing incidents and genuine scheme trade-offs since agents (or elder engineers pairing exclusively alongside agents) assimilate that work, the industry volition endure a critical scheme for expanding the next generation of engineers alongside judgement and taste.

Read that clause again — “senior engineers pairing exclusively alongside agents.” Hold onto it; it’s the accelerant, and I’ll arrive rear to it. First, the device it’s breaking — since to see why this is a genuine shortage and not fair hand-wringing, you have to see what a younger engineer was really for.

Why We’re About to Stop Making Seniors

Step rear and appearance at what hiring a younger engineer actually bought you, for the complete former of this industry. It bought you two things, not one — bundled so tightly that nobody bothered to drag them apart:

1. The activity they did. Real tickets, genuine fixes, genuine features. Not glamorous, but it shipped, and it approximately paid for their salary.
2. The elder they were becoming. Every one of those tickets was additionally a rep. The difficult material — fine application taste, judgement concerning what breaks and why, the gut awareness for whenever item is off — none of it comes from a publish or a bootcamp. It’s learned on the job, slowly, doing genuine activity under person who has already been there.

That second item was the genuine product. Hiring juniors was always a provision sequence for seniors — we fair never had to call it that, since the archetypal item quietly paid for the second. The on-the-job training rode along, basically for free, on top of labor you were purchasing anyway. It looked akin “hiring cheap hands.” It was really “manufacturing seniors, subsidized by the hands.”

Now observe what AI does to that bundle: it splits it spotless in half. The activity — item one — is increasingly done by agents, faster and cheaper than any junior. So the spreadsheet says the apparent thing: stop hiring juniors, the activity is handled. And it’s deceased correct concerning the work. But it silently cancels item two correct alongside alongside it — the OJT pipeline that was the lone scheme we always had for expanding seniors. AI made the junior’s labor redundant, and companies are concerning to throw out the junior’s training value alongside alongside it, since the two continually came in one bundle and nobody always put a row item on the fractional that mattered more.

That’s the entire item in one sentence: we’re concerning to halt making seniors since the item that used to pay for making them fair got automated.

You Can’t Just Buy Them

I can comprehend the objection, since it’s the apparent one and I’ve made it myself: fair employ seniors whenever you need them. Let the another guy train them. Poach the completed product. It’s cheaper, it’s faster, and it has worked for the complete former of this industry. Right? WRONG.

Here’s why that breaks: everyone is concerning to have that exact identical idea at the exact identical time. Every business that quietly decides juniors are too costly is also quietly deciding to buy its seniors on the open market later. There’s indication it’s already starting: hiring at top tech firms is up approximately 20% year-over-year.¹ So in a few years you have the entire industry bidding for a pond of elder engineers that nobody bothered to replenish. Demand goes vertical. Supply is flat — worse than flat, since we all stopped feeding the pipeline at once. You already cognize what that does to a price.

You don’t equal have to ideate it. Look at what AI fair did to memory. Data centers got hungry, petition went vertical, and DRAM prices rose additional than 400% from the commencement of 2024 through the end of 2026 — nearly doubling in a sole quarter at the peak, alongside the deficit forecast to run former 2028 and maybe, per SK Hynix, former 2030. Why so brutal? Because you cannot conjure a fab in a quarter. The capability takes years to build, nobody built adequate of it onward of the demand, and so now everyone pays through the nose for the scarce thing.

Memory fair ran this play. Senior-engineer comp is set up to run it next — flat for years, afterward vertical the instant petition outruns a provision nobody replenished.

Senior engineers are the DRAM of the next few years. Same shape, exactly: petition exploding — all shop on earth wants a elder who can run a fleet of agents alongside judgement — provision seriously constrained, and a guide period you cannot compress by throwing prosperity at it in the moment. It takes years to create one, and no acquisition command speeds that up. When the scarcity hits — and it volition — the companies that kept training through the lean years volition have their own supply, grown at cost. Everyone alternatively volition be bidding against all another for whoever’s left, at any the market demands. And a cornered market demands a lot.

So flip the framing entirely. Growing your own engineers isn’t philanthropy and it isn’t a feel-good morale program. It’s a hedge. You are purchasing senior-engineer futures at today’s price, insuring yourself against a scarcity spike you can already see coming on the tape. I made nearly this exact disagreement concerning buying your own conclusion hardware — expend a small chief now so you’re not at the mercy of a provision you don’t authority later. Talent is the identical bet, and it’s a bigger one. Luck favors the ready.

The Trap That Pours Gas on It

So we concur there’s a deficit coming. Here’s the ruthless part: the way most shops are using AI correct now is accelerating it. Meet the accelerant I told you to clasp onto.

There’s a terrific writeup from the folks at FlintAI — their squad is now 5 devs and 45 agents, all living in Slack. Nine agents per developer. The agents obtain roles — triage, deploy, assessment — the individual steps in for the two decisions that really matter (code review, publish approval), and they study item akin a 3x efficiency bump. It’s a genuinely notable part of engineering and you should peruse it.

That’s the example everybody’s racing toward correct now: obtain your finest seniors, strap nine agents to them, and crop the multiple. And it works! I’m not knocking it (and I’m doing it too)!

But appearance at the form of it. One senior, nine agents, 3x. If you have five seniors, you have five of those. That’s linear. You add leverage exactly as accelerated as you add seniors — and we fair established you can’t create seniors fast, and they’re concerning to get scarce and expensive. You’ve built your complete leverage narrative on the one input that’s turning into DRAM. Worse: all hr that elder spends conducting their personal delegate orchestra is an hr they’re not spending minting the next senior. The nine-agent elder isn’t fair linear. Left alone, it’s decreasing — a fab quietly operating its own provision down.

But Your Business May NEED to Focus Only on Seniors!

What you should do concerning this depends on your business. If you’re a three-person startup clawing for product-market fit before the prosperity runs out — go fast, go alone, strap on the nine agents and run. You don’t have a next generation to develop yet; you have a runway to beat, and that’s fine.

But if you’re a expanding business that plans to motionless be current in five years and to need elder engineers you don’t yet have, the calculus flips hard. You are, whether you’ve named it or not, in the endeavor of manufacturing seniors — and you fair watched the subsidy that funded it disappear.

The Only Hedge: Fan Out Through People

There’s a row you’ve likely heard: “If you desire to go fast, go alone. If you desire to go far, go together.” It’s normally served up as an African proverb but it’s the entire disagreement in nine words. The nine-agent solo elder is going fast, alone. Mentoring is going far, together. Speed and extend are not the identical race.

Now run the another play. Take that identical elder and, alternatively of having them personally run nine agents, have them ALSO mentor three juniors — and instruct each of those juniors to run their own set of agents.

And what should the juniors build? Real, precious activity the endeavor needs — but contained, isolable, testable, the benevolent of item you can clasp in your caput all at once. (I’ll get particular concerning the what and the how additional down.)

Do the arithmetic on the shape, not the exact numbers. One elder immediately produces several output. But a elder who additionally levels up three juniors, all of whom is now driving their own collection of agents, has produced three additional nodes that all fan out on their own — and those juniors rotate into seniors who mentor their own juniors. That’s not one elder times nine agents. That’s a branching tree. It compounds. It’s the difference between a row and a curve, and complete a brace of years the curve isn’t close. You can’t conjure a elder in a fourth — but this is how you interrupt dirt on the fab now, so you have your own provision whenever everyone alternatively is bidding for scraps.

The senior’s most precious output was never the code. In the agentic era it definitely isn’t the code — the agents compose the code. The senior’s most precious output is more group who can create output alongside judgment. That’s the highest-leverage item a elder individual can do now, and it’s the item the nine-agent-conductor example quietly skips.

Left: one senior, nine agents. Linear, and it burns the individual who makes additional people. Right: one elder mentors three juniors who all run their own agents. The senior's genuine output isn't code — it's additional group who create output. That branches.

To Be Clear: Seniors Still Build

Don’t peruse any of this as “stop coding and go be a full-time mentor.” Ugh, no. I wrote a entire article concerning how the bosses are coding again and I meant all term — the seniors entirely motionless need to be building. You cannot mentor judgement you’re not actively practicing. The day you halt shipping is the day your feedback goes stale, your instincts drift, and the juniors can odor it. You can’t study to float from a publish or by observing from the phase for two years — that goes for the trainer too. Get in the water. Stay in it.

The change isn’t whether seniors build. It’s what fraction of their time construction gets. The trap isn’t a elder who codes — it’s a elder who codes alongside 100% of their time, operating their personal nine-agent orchestra and calling it leverage during the place rearward them stays empty. Carve out genuine period — not the scraps, not “if I get to it” — to develop people. Build alongside most of your time; mentor alongside a genuine piece of it, deliberately protected. Both, on purpose. A chief who owns the complete crucial way and mentors no one isn’t scaling; they’re a sole item of nonaccomplishment alongside awesome throughput.

But You Can’t Just Hand Them the Keys

Here’s anywhere it gets hard, and anywhere most “just let the juniors use AI” takes autumn apart. You cannot fan out through juniors who don’t have judgment, since a younger driving nine agents alongside no judgement isn’t a power multiplier — it’s a slop cannon pointed at your manufacturing systems. The FlintAI folks are correct concerning that. Judgment is the entire ballgame.

So the genuine inquiry isn’t “should juniors use agents?” (obviously yes) or equal “should seniors let them?” (yes, and it’s going to awareness bad — additional on that in a second). The genuine inquiry is: how do you build judgement in a individual whenever the tool volition happily do the item before they’ve understood it?

I’ve been chewing on this difficult adequate that I sat downward and wrote an genuine betterment scheme for it. The center understanding is this: agentic tools eliminate the conflict of producing code and infra changes, but they do not eliminate — they increase — the need to verify them. The nonaccomplishment manner of the next few years isn’t juniors who can’t category the commands. It’s juniors who can’t inform whenever a tool’s assured output is locally accurate and globally wrong. The manifest that applies cleanly but misconfigures asset limits. The fix that resolves one node’s indication but ignores the partition. The manuscript that runs successfully and deletes the incorrect thing.

So What Am I Planning?

Concretely, here’s the form of it. First, provision the juniors genuine activity alongside a genuine blast radius, kept deliberately small: canary tests, smaller services, SDKs, example clients that twice as test harnesses. Stuff the endeavor genuinely needs — but contained, isolable, testable, the benevolent of item you can clasp in your caput all at once. They container item that matters and they get the reps. Nobody learns judgement on a toy.

Then drill them. Not metaphorically — literally.

• Chaos drills, not fair tickets. This is the part I’m most enthusiastic about. Take the Chaos Monkey idea and rotate it into training: deliberately interrupt things — slay a node mid-rollout, partition the network, and so on — on a collection that’s harmless to wreck, and put a younger in forefront of it alongside a clock running. It’s a casualty drill direct out of the Rickover playbook, for software. And predict-then-observe is the entire point: before they contact anything, they say out noisy what they anticipate to see and why. The gap between the prediction and what really happens is the instruction — far additional than whether they eventually fix it.
• Certs as a forcing function, not a complete line. I don’t attention concerning the sign on the wall. I attention that a performance-based exam puts a individual solitary at a live terminal alongside a damaged scheme and a clock, which is exactly the confrontation alongside nonaccomplishment that agents document over. Where a dependable cert exists, use it as scaffolding to justify the confrontation happens. I’m preparedness on LFCS, CKA, and AWS SAA, since my juniors live in a Linux/cloud world.
• No-AI-assist checkpoints. At smallest one exercise per accomplishment area alongside the agents turned off, under period pressure. Not since agents are bad — since the lone way to cognize the mental example exists independent of the tool is to obtain the tool distant and watch.

And here’s the part group get exactly backwards, so let me say it loud: use AI to study all of this 10x faster. The identical tool that threatens to skip the apprenticeship is the finest tutor a younger has always had. Stuck on how Raft handles a divided vote? Ask. Can’t see why the NetworkPolicy is silently dropping traffic? Have the example stroll you through it, afterward go verify it on the cluster. Grinding toward the CKA? An delegate volition quiz you, explain all incorrect answer, and create caller broken-cluster scenarios all night long. My juniors volition study faster than I always did — I learned this material from thick O’Reilly books and Usenet threads, and they’ve got a tolerant expert on tap. That’s a gift. Use it hard.

The one rule: you can’t lean on it 100%. Learn with it constantly, afterward demonstrate — at the no-AI checkpoint, under the clock — that the cognition is in your caput and not fair in the environment window. Lean on AI to learn; don’t lean on it to know. Both can be true, and the area between them is the entire ballgame.

This Is Just Rickover, Again

If this appears acquainted to longtime readers, it’s since it’s the identical fine I continually diagram from. The US Navy figured out how to put a 22-year-old in forefront of a live nuclear reactor and rely them not to melt it, and they did it decades before anyone stated “agentic.” The Rickover program didn’t activity since the kids memorized procedures. It worked since it was built, top to bottom, about rising standards, changeless training, predict-then-observe, and escalating bad news fast. You qualified by proving judgement under failure, not by reciting the manual. And all qualified operator’s job included training the next one. The cognition fanned out, by design, or the fleet didn’t have adequate operators. Sound familiar?

Somebody stood near adequate to drag me out of the water. That's the debt. The agents don't cancel it.

I attention concerning this one additional than the tech, honestly. I was displaced as a teenager. I grew up on welfare. I did not claw my way into this occupation on raw deserve solitary — mentors reached rear and pulled, complete and over, and I would not be penning this from a home in the Bay Area if they hadn’t. So whenever I say mentoring the next generation is an obligation, I don’t average it as a nice LinkedIn sentiment. I average I owe a debt I can lone pay forward, and so do you, and the agents don’t get us off the hook. If item they lift the stakes, since the uncomplicated way — hoard the leverage, skip the humans — has never been additional tempting or additional available.

Why Your Seniors Will Hate It

Letting a younger build the item — agentically, their way, at their gait — is slower than doing it yourself, and it’s WAY slower than strapping nine agents to your own elder brain and cranking. You volition observe them go downward a highway you’d have skipped. You volition desire to catch the keyboard. Every intuition sharpened complete your occupation screams just do it yourself, it’ll obtain ten minutes. And you’ll be correct — it would obtain you ten minutes, and it’ll obtain them a day, and at the end of that day you have item additional precious than the feature: you have a individual who’s a notch nearer to not needing you. You fair added a component of the item that’s concerning to be scarcest.

That’s the trade. Short-term speed for long-term capacity. The mentor pays a tax now and compounds later. You can’t study to float from a book, and your juniors can’t study judgement from observing you swim, either — you have to let them get in the h2o and thrash a small during you remain near adequate to drag them out.

Conclusion

The naysayers are correct that the old apprenticeship is dying. They’re incorrect that nothing replaces it. The apprenticeship changes shape: juniors who run agents with taste can study faster than we always did — if person bothers to build the taste. That “if” is the complete post.

So do the math the remainder of the market is concerning to do the costly way. Senior engineers are the next DRAM: petition going vertical, provision nobody replenished, a guide period no acquisition command can compress. The cost is going to spike, and the shops that kept training through the lean years volition have their own provision during everyone alternatively pays ransom for scraps. Growing your own is the hedge — you’re purchasing elder futures at today’s price. The scheme is mentorship: fan out through group alternatively of hoarding agents on the seniors you already have. One elder who makes three who all create three additional is a curve, and curves win. Go solitary and you’ll go fast; go together and you’ll go far. Have powerful opinions, loosely held — but clasp this one tight: the highest-leverage item you can do in the agentic era is still, stubbornly, to build another Engineer, not fair build it yourself faster.

What’s next for me: I’ve got that betterment scheme — four pillars, certs as forcing functions, a distributed-systems practicum alongside no exam since none exists yet — and I’m really operating it, not fair penning concerning it. I’ll study rear on what survives communication alongside reality, honestly, including the parts I get wrong. That’s the deal.

If you’re wrestling alongside the identical item — or if you think I’m romanticizing mentorship and the genuine answer is fair additional agents — autumn me a note on LinkedIn. I particularly desire to comprehend from anyone really expanding juniors in an agent-heavy shop correct now; that’s the test that matters. And yes — if this helped, pay it forward. Somebody did it for me.

 Share!