In 1991, Desert Shield turned into Desert Storm. U.S. Transportation Command had to move people and materiel from Europe into Saudi Arabia on crisis clocks. An AI logistics planner called DART, the Dynamic Analysis and Replanning Tool, went into that fight. Years later, a Computing Research Association overview attributed to DARPA a blunt claim: DART had completely repaid three decades of agency investment in artificial intelligence research.
Treat that as a DARPA-reported claim, not an independent audit. It is still the kind of sentence a field writes when it wants to prove it shipped.
A victory story. Also a late chapter.
Worth rewinding - not for nostalgia, but because the boom that made DART thinkable and the winter that almost buried the field still define how agent stacks fail today. The question underneath every hype cycle is the same: what happens when ambition outruns the stack that has to carry it?
Ambition outruns the stack
Summer 1956. Dartmouth College, Hanover, New Hampshire. John McCarthy - then a young math professor - organizes an eight-week gathering with Marvin Minsky (Harvard), Nathaniel Rochester (IBM), and Claude Shannon (Bell Labs). They ask the Rockefeller Foundation for money. The Foundation awards half of the requested $14,000. It is enough.
They coin a name for the bet: artificial intelligence.
The written conjecture is famous for a reason. The project would proceed on the idea that every aspect of learning or any other feature of intelligence can, in principle, be so precisely described that a machine can be made to simulate it. McCarthy was bullish. He hoped that by summer 1956 he would have a model of such a machine fairly close to the stage of programming in a computer.
About twenty scientists rotate through. Some stay the summer. Some drop in for hours. Conference notes catch the flavor: McCarthy interested in a program that will write programs; Warren McCulloch arguing the brain is a Turing machine. Only a few hundred computers exist in the world. Each is the size of a room. Together they offer a fraction of what a phone holds now. Computers of that era could barely play tic-tac-toe.
The workshop does not build the thinking machine. The name sticks anyway. AI becomes a scientific discipline with a founding myth, a research agenda, and - soon enough - a credibility problem.
General methods were supposed to scale. Symbolic logic. Search. Clever algorithms that would eventually cover every domain. By the late 1960s and 1970s, that hope collides with engineering reality. Funding dries up in patches. Hard ceilings appear on what pure search can do. The field’s first winters are not moral failures. They are invoices: ambition charged against a stack that cannot pay.
Survival means a paradigm shift. Stop hunting for a universal algorithm. Start encoding what specialists already know.
Chemistry first: knowledge becomes the product
Worth recalling why the Knowledge Principle landed as hard as it did - rewind to Stanford in the mid-1960s, before “expert system” was a product category.
Joshua Lederberg wants machines that help with scientific hypothesis formation. Edward Feigenbaum brings the AI craft. Bruce Buchanan helps build the programs. Carl Djerassi and chemists from Stanford’s mass spectrometry world supply the domain. The target problem is concrete: take mass-spectral data for an unknown organic compound and propose plausible molecular structures.
The system that grows out of that work is DENDRAL. Its empirical lesson is not “search harder.” It is that competence comes from specialized chemical knowledge packed into the machine - heuristics chemists actually use - more than from a general problem-solver wandering a huge hypothesis space.
Knowledge engineering gets a name in that transfer. A running program that is “stupid” in a domain expert’s eyes becomes useful the moment you ask the right question: what do you know about this problem that the program does not know? Then you encode the answer. Rule by rule. Interview by interview.
That is the turn. AI stops being only an academic bet about mind-in-general. It becomes a craft for extracting and operationalizing expertise.
Douglas Lenat and Edward Feigenbaum later state the lesson as the Knowledge Principle: a system exhibits intelligent understanding and action at a high level of competence primarily because of the specific knowledge it contains about its domain of endeavor. General reasoning processes, being general, are weak. In the absence of knowledge, reasoning will not carry the load.
If you remember one sentence from the expert-systems era, make it that one. Everything that follows - glory, winter, statistical revival, RAG, agents - is a fight over how knowledge gets into the machine and what happens when it doesn’t.
MYCIN: proof of method, failure of deployment
MYCIN grows out of the same Stanford Heuristic Programming Project. Edward Shortliffe’s doctoral work in the early 1970s builds a rule-based system for serious bacterial infections - bacteremia, meningitis - recommending antibiotics with dose adjusted for body weight. It runs in Lisp on a DEC PDP-10 time-sharing system reachable over the early ARPANET. It can explain its reasoning. It introduces certainty factors - numeric confidence scores attached to each rule (how strongly the evidence supports a diagnosis) - so the engine can weigh uncertain medical evidence instead of treating every fact as absolute.
It is never used in routine patient care.
Not because the research was empty. Observers raised ethical and legal questions about responsibility when a computer is wrong. The larger barrier was integration. MYCIN was a stand-alone consultation: a physician typed answers to questions; nothing plugged into hospital records. By the late 1970s the research system was effectively static. Within a few years its knowledge base aged in public - new antimicrobial classes arrived, and a frozen rule set could not keep up without maintenance machinery the project did not have as a clinical product.
MYCIN’s lasting influence is methodological. Rule-based shells spread. EMYCIN extracts the inference machinery from the medical content. And one difficulty rises to a name that will haunt the boom: the knowledge-acquisition bottleneck - getting what experts know out of their heads and into rules the machine can fire.
Hold that problem. It does not die with expert systems. It changes costume.
Glory: XCON configures the machine that ships
Before XCON, ordering a VAX from Digital Equipment Corporation was a parts explosion. Every cable, connection, and software piece was a separate line item. Computers were not sold as sealed boxes. Salespeople were not always deep technical experts. Customers received hardware without the right cables, printers without the right drivers, processors without the right language chips. Delays. Dissatisfaction. Legal action.
John McDermott at Carnegie Mellon builds R1 - DEC calls it XCON, Expert Configurer - in OPS5, a production-rule language. Work starts in 1978. Experts from DEC’s PDP/11 and VAX groups disagree with each other about optimal configurations; sorting that out becomes part of the product quality. The system goes into regular manufacturing use in January 1980 at DEC’s plant in Salem, New Hampshire. It asks critical questions, then prints a coherent order.
A typical system is fifty to a hundred and fifty components with spatial and logical constraints among them. XCON grows to about 2,500 rules. By 1986 it has processed roughly 80,000 orders at 95–98% accuracy. Estimates put savings around $25 million a year - fewer free replacement parts after configuration mistakes, faster assembly, fewer angry customers.
Bachant and McDermott, writing after four years in the trenches, report that fewer than one in a thousand orders is misconfigured because of missing or incorrect rules - and they are honest about why gaps remain: even after tens of thousands of orders, the system has seen only a fraction of possible situations, and new products sometimes ship before the rule base catches up.
Success breeds siblings. XSEL helps the salesforce configure so a customer does not order a machine that will not fit through a doorway or choose too few cabinets for the boards. XSITE handles site and location constraints.
This is peak industrial symbolic AI. Real throughput. Real money. Knowledge Principle as a P&L line.
And now - six years after Salem, while XCON is still earning - the wider market for “AI as product category” is already cracking.
Winter: four failure modes, one moonshot
Expert systems do not die because one paper refutes them. They die as a cascade of engineering and economic failures that investors eventually refuse to refinance.
Knowledge-acquisition bottleneck. Specialists often cannot turn tacit judgment into IF-THEN rules. The decisions that look like genius in a human are the ones hardest to interview into a knowledge base. Every hour of elicitation is expensive. Every missing rule is a future outage.
Brittleness. Outside the ontology - the fixed set of categories and facts the rule base was built to recognize - systems do not degrade gracefully. They fail hard instead of saying “I don’t know.” Humans soft-fail. Rule bases cliff-edge.
Complexity catastrophe. As rule counts climb, interactions become unpredictable. Change one principle; break something distant. XCON’s own maintainers lived the milder version of this - continuous knowledge-base work as the product line moved. Scale that across many commercial systems with weaker engineering discipline, and the industry’s reputation burns.
Hardware economics. Dedicated Lisp machines lose to cheaper general-purpose workstations and PCs that close the performance gap. The specialized box stops being a moat. It becomes a stranded asset.
Japan’s Fifth Generation Computer Systems project (1982–1992) is the clean moonshot case. Japan’s Ministry of International Trade and Industry funds a decade-long bet on massively parallel, logic-oriented machines - Prolog-flavored architectures meant to host a new AI era. Contemporary American coverage framed it as Japan’s computer challenge to the world; Feigenbaum’s 1983 book The Fifth Generation was part of that framing. Research comes out. Concurrent logic programming advances. A commercial platform that displaces general-purpose computing does not. When the project closes in 1992, the verdict from market reality is blunt: specialized parallel hardware loses to Sun workstations and Intel machines the same way Lisp machines lose to ordinary PCs running rule engines like CLIPS.
In the U.S., systems like XCON and later DART at least get forced through real deployments - manufacturing floors, logistics commands, invoices and after-action reports. Isolation from that feedback loop makes a national moonshot harder to correct mid-flight.
The winter is not “AI was fake.” The winter is “ambition sold ahead of maintainable knowledge, graceful failure, and boring hardware economics.”
Remember DART’s victory sentence from the cold open. It lands after this winter. Quiet progress, then a logistics tool in Desert Storm, then a claim about thirty years of investment. That rhythm - overpromise, freeze, underground engineering, rebirth - is the cycle. It will repeat on a statistical substrate.
Two traditions: neat rules, scruffy weights
By the early 1990s the field has polarized. Camp names vary. The trade-offs do not.
Symbolic tradition. Explicit knowledge. Logic. Inspectable decision paths. You can audit why the machine said no. When the world stays inside the model’s fixed categories, that is power. When it leaves, you get brittleness.
Statistical tradition. Learn patterns from data. Neural nets. Support-vector machines. Later deep learning. Noise tolerance. Generalization. When the world is messy, that is power. When you need to know why, or when the weights invent a fluent falsehood, you get a different kind of winter - trust erosion instead of rule-base collapse.
On the symbolic side, Judea Pearl’s work on Bayesian networks through the 1980s gives AI a mathematical language for reasoning under uncertainty without pretending the world is a pile of crisp IF-THEN facts. Probability becomes infrastructure, not a footnote.
On the connectionist side - the statistical, weight-learning camp - John Hopfield’s 1982 work shows how ideas from statistical physics - energy minimization - can describe associative memory in networks. Then 1986: David Rumelhart, Geoffrey Hinton, and Ronald Williams publish on learning representations by back-propagating errors. Multilayer nets can learn internal features again. The connectionist research program, which had been out of fashion, gets a workable training algorithm and a generation of students.
Support-vector machines bridge the statistical story into the 1990s commercial toolkit. Vapnik’s statistical learning theory builds the foundations; Cortes and Vapnik’s 1995 soft-margin SVMs make the method practical when training data is not cleanly separable. For a stretch before deep learning takes the crown, SVMs are a default for high-dimensional classification - especially when features outnumber samples.
Worth pausing on the next substrate shift - without it, the foundation-model boom later in this piece has no machine underneath it.
In 2017, eight researchers at Google - Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin - publish Attention Is All You Need. The paper proposes the Transformer: a sequence model built entirely on attention mechanisms, dispensing with recurrence and convolutions. On the WMT 2014 English-to-German and English-to-French translation benchmarks it beats the then-dominant recurrent and convolutional stacks while training faster. Google’s own write-up puts the speedup at up to an order of magnitude - a better fit for GPUs and TPUs, which excel at parallel work rather than left-to-right steps.
That is the hardware-economics rhyme again. Lisp machines lost to ordinary PCs. Sequential RNNs lose ground to attention that can saturate a GPU. Parallel training is what lets the statistical tradition scale into the mass-market generative systems the next sections will meet - BERT, GPT-class models, and the foundation-model wrappers that agents sit on top of.
The split never fully closes. It just changes what the models run on. Every decade re-asks the same hybrid question: how do you keep the flexibility of learning without surrendering the control of knowledge?
Exhibit: timeline
| Year | Event / system | What it established |
|---|---|---|
| 1956 | Dartmouth Summer Research Project | Name and founding frame for AI as a field |
| 1960s | DENDRAL | Domain knowledge beats bare search for scientific hypothesis formation |
| 1969 | Frame problem (McCarthy & Hayes) | Formal difficulty of stating what stays true after an action |
| 1972–1976 | MYCIN | Rule-based medical diagnosis as a research landmark; deployment blocked by integration |
| 1978–1980 | XCON / R1 | Industrial-scale rule configuration at DEC |
| 1982 | Fifth Generation (Japan) | National moonshot on parallel logic machines |
| 1982 | Hopfield networks | Statistical physics applied to associative memory |
| 1985 | CLIPS (NASA) | Portable C rule engine; expert systems on ordinary PCs |
| 1986 | Back-propagation (Rumelhart, Hinton, Williams) | Multilayer nets learn internal representations |
| Late 1980s | Judea Pearl / Bayesian networks | Probabilistic inference under uncertainty |
| 1991 | DART | Logistics AI; DARPA-reported ROI claim covering 30 years of AI investment |
| 1995 | Soft-margin SVM (Cortes & Vapnik) | Pre-deep-learning standard for high-dimensional classification |
| 2017 | Transformer (Vaswani et al.) | Attention-only architecture; parallelizable substrate for large language models |
| 2020 | RAG (Lewis et al.) | Parametric generation plus non-parametric retrieval memory |
| 2022– | Foundation-model boom | Mass-market generative AI |
| 2024– | Agentic systems | Models wrapped in tools, plans, and harnesses that act |
Exhibit: why SVM still shows up
That same split shows up in tooling. Support-vector machines still earn a seat where data is structural and scarce - the classic high-dimension, low-sample regime. Kernel methods map data into richer feature spaces without training a deep net that overfits a tiny set. Niche lesson from the winters: the method that survives matches the data you have, not the loudest demo.
Exhibit: the players
| Role | Who / what | Why they matter here |
|---|---|---|
| Founding frame | McCarthy, Minsky, Rochester, Shannon | Name the field; bet on describable intelligence |
| Knowledge Principle | Feigenbaum, Lenat; DENDRAL team | Domain knowledge over bare search |
| Medical rules | Shortliffe / MYCIN | Method works; integration and maintenance decide deployment |
| Industrial proof | McDermott / XCON at DEC | Rules on a P&L; then the maintenance bill |
| Moonshot caution | MITI Fifth Generation | Specialized hardware loses to general-purpose economics |
| Statistical reopen | Hopfield; Rumelhart, Hinton, Williams; Pearl | Nets and probability after the symbolic peak |
| Scaling substrate | Vaswani et al. / Transformer | Self-attention that fits GPU economics; backbone of foundation models |
| Hybrid reopen | Lewis et al. RAG; agent harnesses | External knowledge + learned models again |
The hybrid comes back wearing new clothes
Retrieval-augmented generation, as Patrick Lewis and co-authors frame it in 2020, is not a nostalgia project. It is a hybrid architecture with a plain claim: large parametric models store a lot of knowledge in weights, but they cannot easily expand or revise that memory, cannot straightforwardly show their working, and may produce fluent falsehoods - hallucinations. Combine parametric memory with non-parametric memory - a retrievable document index you can update - and you get generations that are more factual, more specific, and more inspectable than weights alone.
Read that against the Knowledge Principle: external documents are a modern way to put domain knowledge next to a reasoner. Stack it against MYCIN’s failure mode: integration and maintenance still decide whether the knowledge stays true. Stack it against XCON’s success: the value appears when the system ships into a real workflow - not when the demo looks clever.
Agentic systems push the same reunion further. Planning, tools, sandboxes, retrieval, and evaluation sit around a statistical core the way rule engines once sat around search. The 1980s ambition - machines that act in a domain on behalf of a user - is back. The substrate changed from OPS5 and Lisp machines to Transformer-based foundation models wrapped in tools and connectors. The failure modes did not vanish. They rhymed.
Knowledge still has to enter somehow. If your “rules” are a bloated instruction file nobody maintains, you have rebuilt the acquisition bottleneck in markdown.
Brittleness still appears at the edges. An agent that cannot say “I don’t know” and stop will improvise - the statistical version of stepping outside the categories it was built to handle.
Complexity still grows with every tool, skill, and subagent you bolt on. Change one hook; break a distant path. The catastrophe is quieter than a multi-thousand-rule XCON, and just as real.
Hardware economics changed shape. The expensive specialty box is less often a Lisp machine now and more often an unexamined bill for tokens, context, and always-on tools. Same lesson: general-purpose efficiency eventually eats romantic architecture.
Compute alone does not buy trust. Organizations that dodge the next winter treat verification and the agent harness as first-class work - not decoration on a bigger model.
Exhibit: sources
Exit: same agent, different harness
The live question is not whether agents work. DART “worked.” XCON “worked.” MYCIN “worked” in the lab and still never reached the ward. The live question is whether you put verification, maintainable knowledge, and honest failure modes around the system before the boom sells the demo.
Same agent, different harness. That is the practical exit from this history: vibe coding vs agentic engineering.