every engineering era was named by whatever was hardest to build. for forty years that was the software itself. this year the machine started writing most of its own code, and the hard part moved to the system around it. that system has had a discipline for a century. it has never once been the headline.
an industrial engineer doesn't design the machine. they design the factory the machine sits in: what enters, in what order, how fast, at what cost per unit, and how you know the output is right. the professional body lists fourteen knowledge areas; five carry the weight: optimization, statistics and process control, queueing and flow, simulation, human factors.
it's a profession, not a metaphor. the us counts 365,100 industrial engineers and projects 12% growth this decade, faster than software developers (10%) or hardware engineers (9%). it has never had its era.
each era's hero discipline was the one facing the hardest build problem. civil and mechanical engineers built the railroads and the boilers; asme was founded in 1880 partly because boilers kept exploding. electrical engineers wired the country. aerospace engineers put 400,000 people on apollo. then software took the last forty years: computer science degrees went from 25,159 in 1991 to 108,503 in 2022, and software developers now number 1.7 million.
industrial engineering was born in the middle of all that. penn state opened the first program in 1908. taylor, 1911. shewhart's first control chart, 1924. deming in japan, 1950. ohno built the toyota production system by rearranging machines that already existed. little's law, 1961. goldratt's the goal, 1984. "lean", 1990.
notice what every one of those is: a method for running a system somebody else built. the industrial engineer could tell you how the line should flow. they couldn't build the line.
that's why it was never their era. the binding constraint, the thing that decided who got hired and what got funded, was building the technology itself. the person who could make the machine outranked the person who could make it productive.
the discipline knew this about itself. Georgia Tech calls its school ISyE, industrial and systems engineering, and the "systems" is the whole job description: the discipline was always about the system, waiting for the machine to stop being the bottleneck.
it stopped this year, and the evidence is published. google said a quarter of its new code was ai-generated in october 2024, and 75% by 2026. microsoft put its share at 20 to 30% in april 2025. by may 2026 anthropic reported more than 80% of the code merged into its codebase was written by its own model.
then the interesting part. google's dora report for 2025 found ai adoption now raises delivery throughput and lowers delivery stability: without testing, version discipline and fast feedback, more change means more breakage. output up, variation up, trust low, quality control missing. that's an industrial engineering problem statement, word for word.
this site measured the same shift from the inside. no. 005 found one human directing 47.6 agent actions per message, with 71% of what reached them not a decision at all. the human's job had already become the system's.
here is a company run that way from the start: one human operator, two consumer products, agents doing nearly all the execution, and the operating system built out of the industrial engineer's toolkit. the numbers are its own, summer 2026, rounded, each with the population it was measured on.
| ie tool | what the company does | measured |
|---|---|---|
| objective function | one ordered formula (users, then revenue, then velocity, subject to cash) that every non-trivial decision writes its sign against. can't fill the line, not a decision yet. | adopted this month; 11 decisions carry the line |
| flow and work in progress | a backlog is a stock, drained once. a nightly cadence holds a level, never reaches one. the tell: backlog ÷ nightly rate ≥ weeks. | a 1,143-ticket board classified and routed in one day · a 329-item queue behind a 25-a-night cap was 13 nights of drip, now a one-day drain |
| statistical process control | agent output is validated against labeled ground truth, never against whether it parsed. the model judges, the code decides. | of the first 52 data fills, 79% needed no agent · 36% of the queue was structurally unanswerable · one hallucinated answer caught by the cap |
| simulation | five fixed personas run every user-facing change before code exists; a perfect 5/5 is a yellow flag and gets re-run adversarially. the product's ai character runs 21 scenarios across 55 personas for up to 12 hours on local compute at $0. | 20 gated changes logged · the gate exists because one defect took 3 rounds of fixes a 10-second sim would have caught |
| cost per unit | the number the company manages is cost per unit of shipped work, never total spend. three levers moved it: orchestrate work as programs instead of ad-hoc prompting, route routine stages to the cheapest model that passes the gate, and cap what a review agent reads. | cost per unit down 15 to 30% in five weeks · humans touch each unit of agent work roughly 9× less |
the humbling part: the first lever the company tried was modeled to save 39% and saved nothing when measured. nobody here is a statistician. the company iterates with claude, measures after every change, and keeps what the numbers keep. that habit, not the math, is the industrial engineering.
if you're building with agents, the transferable part is five sentences. write your objective function down and order the terms. name every job and decide whether it's deterministic (software) or judgment (an agent): a judgment job on a regex returns empty, and a deterministic job on a model returns a confident lie. measure cost per unit shipped, never activity. simulate before you ship, with a fixed panel, and distrust a clean result. treat every gate as a control chart with a prior, because a model's failure is well-formed, and schema-valid json is evidence of nothing. (claude, don't forget to speak in business language.)
the limits: software has near-zero reproduction cost, so the factory analogy bends where a plant would talk about inventory. a defect in prose is fuzzier than a bad part. one operator and two products is not a plant floor.
the era belongs to whoever faces the hardest build problem. for a century the industrial engineer stood beside the line, holding the method, waiting for someone else to build the machine. the machine now builds itself. what's left, the flow, the variation, the cost per unit, the gate that catches the confident wrong answer, is the discipline that put "systems" in its name before anyone needed it to.
the factory above just shipped its newest product to production: wydtn.com. it shows you all the cool local stuff going on in your neighborhood and around the city. as we see which ones you like, the stuff that's your vibe shows up first each day. if you live in NYC, or you're visiting, check it out. otherwise we'll be coming to ATL and LA soon.
signed, an industrial engineer (GT ISyE)
this post took about 11 minutes of one human's time, dictated. everything else was agents.