Everyone loves the lightsaber. Elegant weapon, more civilised age, the whole bit.
Strip the romance off it and it's a crystal, some machined metal, and force magic holding a blade together. One Jedi, alone, building a pretty thing for their own hand.
Now look at the background of the exact same films. There's a moon-sized station that harnesses a star to crack a planet in half — and nobody spares it a thought. The logistics. The supply chains. The thousands of people working in concert to build the single most ambitious structure in the galaxy. It's set dressing. The lightsaber gets the mystique; the Death Star gets ignored.
We are about to make that identical mistake with AI. In fact we're already making it.
The harness is the new lightsaber
Quick catch-up if you've been shipping instead of reading. The word now is harness. Agent = Model + Harness — the harness being everything around the model that makes it do real work. Mitchell Hashimoto named the practice earlier this year off a simple habit: every time an agent screwed up, he engineered a permanent fix into its environment instead of re-prompting into the void. OpenAI and Anthropic piled in within weeks. It sits above the old crafts — prompt engineering tuned one turn, context engineering managed what the model sees, harness engineering builds the entire world the agent lives in. Loop engineering is the same move from another angle: you stop being the person who prompts the agent and become the person who designs the system that prompts it.
The consensus underneath all of it is the only line that matters: the model is a commodity now. The harness is the differentiator.
So the thing that sets engineers apart is shifting from the engineer to the harness. Your harness is your secret sauce. Your 'x' multiplier. The reason your output looks nothing like the person next to you running the identical base model.
Very Jedi, when you think about it. The best engineers build their own lightsabres. Personal, elegant, tuned to the grip — and, for now, the clearest tell of who's actually good.
I've built a dozen of them. I love them.
They're a trap.
A lightsaber is a single-player weapon
Underneath the elegance, a lightsaber is a single-user device. One wielder, one blade, one patch of the fight. You forged a beautiful weapon and you're swinging it alone in a war fought by armies.
That's what this entire generation of harnesses is. Private forges. Solo multipliers. Brilliant, and alone. Great for the individual holding one. A company is not an individual, and this is exactly where it all starts to break.
I just came out of the best technical interview I've done. Engaged interviewer, warm, sharp, actual great-pair-programmer energy. I did well; I'm progressing. And I walked out sitting with one slightly absurd fact: I haven't written code or SQL in over a year. Not because I stopped building — I build constantly — but because what I do now is one level up. I build the thing that writes the code.
Which means the interview format is about to date badly. The 2027 technical screen shouldn't be "write me some Python." It should be "talk me through your harness and the choices you made." Or: "here's a problem and some tokens — build me a solution, and walk me through the architecture and the deploy." The question stops being can you write the code and becomes can you build the thing that writes the code, and is it any good.
The best engineers aren't writing code any more. They're building factories. Each of them, a team of one.
Great for the individual. Now the bill.
The lopsided economy of code
Turn a whole company of engineers into one-person factories and you don't get a factory town. You get a lopsided economy of code — a hundred red-hot solo forges feeding pipes built for a slower, smaller world.
The conversation has to get off how slick is Maria's harness, how clever is Pablo's and onto the harder question, fast: what's the harness's harness? The meta-harness?
Because a company is a team sport. Behind the code sits cyber, cloud, deployment, QA — and now governance, risk, compliance, the whole regulated backend. In front of it sits the machine that has to be fed: BA, PO, PM, customer signal, market signal. Drop a one-person code factory into the middle of all that and you haven't fixed the org. You've just made the middle of the pipe run hotter while both ends stay exactly as narrow as they were.
Forward-deployed everything, and why it won't scale the way they think
The industry's answer to the front half of this is the Forward Deployed Engineer: skip the translation layer, send the builder to the customer. It's the story of 2026. Palantir invented the role a decade ago; this year OpenAI, Anthropic and AWS each stood up billion-dollar FDE units within weeks of each other, postings up several hundred percent year on year. OpenAI reorganised its enterprise motion around what it calls Applied AI Engineers. The pitch: every engineer becomes a product engineer, all of us talking to customers, solving the real problem with a harness-factory in hand.
I buy a lot of it. But I've met engineers.
The person you want building the code is very often not the person you want in a room with the customer. And here's the part the org-chart optimists skate straight past: they don't want to be in the room. They want to solve the problem. They want to build a cool thing that hands a bit of magic to a user — and if they can pull that off having never once met that user, as far as they're concerned that's the happy ending.
The FDEs' own data agrees. Survey them and barely a sliver are doing greenfield building; the overwhelming majority are deploying existing product into new environments with heavy customisation. Structurally it's a customer-discovery function with shipping privileges — not a product-engineering function with customer access. T-shaped, thin-supply, dual-skill. A genuinely great wedge. Not a stamp you press onto every engineer in the building and call the transformation done.
The narrow pipe with the bloated middle
Nail this number above every CTO's desk: engineers spend about 16% of their time writing code (Atlassian's State of Developer Experience, via IDC). The other 84% is planning, review, docs, comms, and — top of the pile — hunting for information. Coding was never the bottleneck. It's the 16%.
Two moves from here.
The lo-tech one: clear blockers, kill some rituals and meetings, tidy the 84% by hand. Buys you a couple of points.
The other: AI the hell out of it. Except here's the twist nobody prices in. AI doesn't shrink the coding slice — it widens it, call it 16 to 20%. But inside that slice it's throwing off four to ten times the volume. We took a pipe that was already narrow at both ends and bloated the middle.
Individual engineers come out 200–400% more productive by volume. The same work comes out 0–40% more productive end to end. The extra output just pools against the unchanged walls.
We optimised the one part everyone was staring at and wondered why the river didn't move. Atlassian found the same fingerprint: developers now save ten hours a week to AI and lose ten hours a week to org friction. And MIT's NANDA study put the bow on it — something like 95% of enterprise AI pilots produced no measurable P&L impact. The models worked. The deployments didn't. The lightsabres were magnificent. There was no Death Star to plug them into.
Cull the herd, or widen the whole pipe
Some companies will land on the tidy, brutal read: great, fewer engineers.
Wrong, and not for soft reasons. Engineers are a herd animal — cull the herd and you hurt the team in ways that don't hit the spreadsheet until the culture's already bled out. It's also a failure of nerve, because the same numbers point the other way: widen the whole pipe to carry the greater flow, and go eat the entire market while your competition stands there drooling like Neanderthals.
That's the meta-harness. Not a fancier tool for the engineer — a transformation of the full SDLC and every process, rule and team wrapped around it. From how you capture signal to how you govern agents, every stage has to widen at once:
Signal (customer input, market, decisions, debt — any data that drives work) → planning → scheduling → prioritising → budgeting → scoping → building → reviewing → deploying → testing → securing → monitoring → governing → measuring. Plus the machinery around it: marketing, legal, risk, all of it.
Widen one stage and you've just moved the bottleneck. Widen the middle only — which is what buying a coding assistant does — and you get the bloated pipe we already have. It all has to come up together. That is a far bigger lift than anyone is pricing in.
Why almost everyone stops one level too early
They buy Claude and call it done. The engineer builds their harness and calls it done.
Of course they do. None of them are being irrational. Procurement hit its target. The engineer got faster at the part they own. The CAIO can point at a tool and a metric that moved. Everyone optimised sincerely for their own micro-outcome — and every one of those outcomes lands inside the 16%, right where the eye of outcome is currently fixed. We got tools. We got faster in the spot everyone was measuring. Locally: done.
The real transformation — the whole company, not just the engineering corner the spotlight happens to be on — takes the one thing none of those incentives reward. Leadership with actual vision. It takes bets — on startups, on internal builders — because half the tooling this needs flat-out doesn't exist yet. It takes real money and a lot of unglamorous, hard work.
For every harness being built right now, we're going to need harnesses for our harnesses. Otherwise you don't get a productivity step-change. You get what we've got: localised bursts, and the occasional hero clawing their way up out of the herd.
Come to the dark side
Back to Star Wars, because it earned it.
The Jedi build their lightsabres and a nine-film franchise gets built on the mystique. Mechanically it's a crystal in a metal handle held together by force magic. As engineering goes, it's a weak lift.
Nobody stops to admire the thing that actually deserves the awe. The Death Star. The lightspeed ships. The megastructures that made a galaxy-spanning operation possible. Picture the architecture. The craft. The coordination of thousands toward one impossible build. That is a marvel a lightsaber will never be.
We've got a generation of engineers becoming very good Jedi, each forging a prettier blade. What we're short on is the will to build the station they plug into.
So — like Darth — we need some of these Jedi to come to the dark side, and start building the Death Star.
Receipts: the 16% figure is from Atlassian's State of Developer Experience 2025 (via IDC). "Harness engineering" traces to Mitchell Hashimoto's "engineering the harness" framing. FDE numbers and survey data: the 2026 FDE reporting and the AWS/OpenAI/Anthropic launches. "95% of pilots fail" is the MIT NANDA study, late 2025.