A history of the pursuit of autonomous machine intelligence, from alchemy to LLM agents
This is a companion reference document, not a transcript or summary of any single talk. It stands on its own: no prior session, slide deck, or course is assumed. It traces one recurring pattern across roughly 250 years of intellectual history — a confident claim about achieving genuine machine autonomy, followed by real and lasting tools that claim actually produced, followed by a quiet retreat from the original claim.
Each section below covers one era. Every section opens with a short paragraph describing that era — its beginning, its middle, its end, and what it left behind. Below that are links, organized in two tiers:
This document is licensed under Creative Commons BY 4.0. Share it, adapt it, teach with it — just credit the source.
Alchemy
↓
Cybernetics
↓
GOFAI
↓
Two Responses
├─ Robotics
└─ BDI
↓
LLM AgentsLong before anyone tried to build a thinking machine, a different discipline spent centuries pursuing a different impossible transformation: turning base metals into gold. For a long time, historians treated this as embarrassing pseudoscience that chemistry eventually grew out of. That view has changed. Modern historians of science now argue that “chymistry” — the older, unified term covering both alchemy and what we’d now call chemistry — was a rigorous, internally consistent field, pursued with real experimental discipline. The goal of transmutation was never dramatically renounced in a single moment. It was quietly set aside as the real payoff — the actual chemistry that alchemical experimentation produced — took over. Mineral acids, distillation, and systematic laboratory method all trace back to this pursuit. The “failure” produced the field that succeeded it.
Start here
Go deeper
In 1943, a mathematician, a physiologist, and an engineer proposed something radical for the time: that purposeful, goal-directed behavior could be explained mechanically — through feedback — without appeal to anything mystical about “will” or “purpose.” A machine that compares its current state to a target and adjusts, they argued, is already behaving purposefully. W. Ross Ashby pushed the idea further with his “Homeostat,” a device that could adapt to disturbances without any internal map of its environment at all — pure trial-and-error stability-seeking. The field that grew out of this, cybernetics, didn’t scale into the general theory of adaptive intelligence its founders may have hoped for. But it didn’t disappear either — it became a component technology, quietly present in every thermostat, autopilot, and cruise control system built since. The word “cybernetics” itself faded from use faster than the idea did.
The era’s clearest popular echo isn’t a specific machine but a rulebook: Isaac Asimov’s Three Laws of Robotics, introduced in 1942, a year before Rosenblueth, Wiener, and Bigelow’s founding cybernetics paper. The Laws aren’t a technical anticipation of Ashby’s feedback concepts — Asimov’s robots follow explicit, ranked rules, which is a different idea from rule-free adaptation. But they capture the same underlying cultural instinct: a machine’s behavior needs to be governed, checked, and corrected. Asimov spent decades of stories showing how faithfully-followed rules could still produce strange, unintended outcomes — an early fictional version of a pattern this document traces across every era that follows.
Start here
Go deeper
In 1950, Alan Turing had posed a deceptively simple question: if a machine’s conversation were indistinguishable from a human’s, what grounds would remain for denying it could think? By the mid-1950s, a new and more ambitious claim had taken hold, offering a route toward exactly that kind of performance: intelligence itself is symbol manipulation. Give a machine the right symbols and the right rules for combining them, the thinking went, and general intelligent behavior would follow. Shakey, a robot built at Stanford Research Institute between 1966 and 1972, embodied this claim directly — it sensed its environment, built a plan using a system called STRIPS, and acted on that plan. A related claim, “knowledge is power,” produced expert systems like MYCIN, a program that diagnosed bacterial infections about as well as a human specialist. Both threads produced real, durable technology — search algorithms and rule-based reasoning still sit underneath route planners, business rule engines, and countless automated decision systems today.
But the underlying claim ran into a wall philosophers now call the Frame Problem: a system reasoning in pure logic has no natural way to know what doesn’t change after an action, and specifying it explicitly balloons out of control. This turned out to be the theoretical face of a much broader pattern. These systems were precise but shallow — they handled exactly what had been explicitly specified, and nothing else. In practice, that meant brittleness, an absence of common sense, and persistent trouble with ordinary ambiguity, nowhere more visibly than in natural language, where decades of dedicated effort never obviously closed the gap to human-level understanding. The era’s most famous public triumphs prove the point rather than disprove it: Deep Blue’s 1997 chess victory over Garry Kasparov and IBM Watson’s 2011 Jeopardy! win were real engineering achievements, confined to narrow, tightly-bounded domains, and neither generalized into anything resembling common-sense reasoning. In 1973, the UK’s Lighthill Report delivered a harsh public verdict on the field’s progress, heralding the first “AI winter” — a period of sharply reduced funding and interest that would recur again in later decades.
Philosopher John Searle supplied the era’s most influential theoretical objection. His Chinese Room argument (1980) imagined a person who doesn’t speak Chinese, alone in a room, manipulating Chinese symbols using only a rulebook, producing perfect answers without understanding a single word. His claim: that’s what any symbol-manipulating computer is doing, no matter how good the rules get. It wasn’t an argument that GOFAI had failed to achieve understanding yet — it was an argument that no amount of engineering along this path could succeed, a categorical objection rather than a practical one.
Start here
As one of the longest eras in this document (two and a half decades, much longer than any of the others) there are naturally more cultural touchpoints here:
Go deeper
Rodney Brooks proposed something that sounded almost like a retreat: stop trying to build an internal model of the world at all. His claim, made in 1986, was that representation itself was “the wrong unit of abstraction” for real-time intelligent behavior — that a robot built from layers of simple, reactive behaviors, each directly wired from sensing to action, could produce useful intelligence without ever building a symbolic model of anything. “The world is its own best model,” as the slogan went. This approach produced real, shipped engineering: reactive control is the entire design philosophy behind the Roomba, iRobot’s household vacuum, released in 2002 — sixteen years after Brooks’ original claim. The Roomba is also, in a strange way, its own best illustration of the approach’s limits: its bump-and-try-again cleaning pattern is charming and effective, but it visibly isn’t planning anything. Critics like David Kirsh argued that some tasks — long-term planning, abstraction — genuinely need internal models, and the field mostly settled into hybrid architectures that combine both approaches. Separately, and much later, iRobot itself collapsed as a company — bankrupt by December 2025, sold to a Chinese manufacturer — though that outcome had far more to do with tariffs, a blocked Amazon acquisition, and competition than with any failure of the underlying idea.
Start here
Go deeper
Around the same years Brooks was proposing to abandon internal models, a different group of researchers proposed keeping some structure — just not full “knowledge” of the world. Drawing on philosopher Michael Bratman’s 1987 account of practical reasoning, Anand Rao and Michael Georgeff formalized the Belief-Desire-Intention (BDI) model: an agent maintains beliefs (what it takes to be true), desires (what it wants), and intentions (what it’s currently committed to doing), and it doesn’t need to constantly re-examine every goal from scratch — it commits to intentions and only reconsiders them under the right conditions. This gave the field a genuine, lasting formal vocabulary for goal commitment, still present today as infrastructure in multi-agent software, game AI, and automated negotiation systems. But BDI itself, as a named research program, seems to have peaked: publication counts for the specific term rose steadily from the late 1980s, peaked around 2013–2017, and have been declining since — even as the broader term “multi-agent systems” kept climbing, propelled by an entirely different, much later wave of LLM-based agent research. BDI didn’t lose an argument. It got quietly absorbed, the same way feedback control was absorbed into control engineering decades earlier.
Start here
Go deeper
A note on this document’s own research: the publication-count claim above (rise, peak around 2013–2017, decline since, contrasted with the still-climbing broader term “multi-agent systems”) comes from original research conducted for this project, not from an external source. See the appendix at the end of this document for the method and full data.
The most recent chapter restates the same underlying claim in new vocabulary: “agentic AI” — models that can reason, call external tools, take multi-step actions, and pursue a goal with some independence, rather than simply answering a single prompt. The ReAct pattern, described in a 2022 paper, interleaves reasoning and action in a loop. By 2023, systems like AutoGPT and the broader adoption of function-calling made this kind of multi-step, tool-using behavior culturally visible for the first time to a general audience. The technique is genuinely useful — bounded, tool-assisted, multi-step execution works, often impressively, within a well-defined task and toolset. But the walk-back is already visible and is, in a real sense, the Frame Problem wearing new clothes: hallucination, brittle tool use, and weak long-horizon planning. Researcher Subbarao Kambhampati and colleagues have argued that large language models, by themselves, are not reliable planners or self-verifiers — they’re best understood as approximate knowledge sources that need external checking, not autonomous reasoners. In 2023, a New York lawyer submitted a legal brief containing fabricated case citations generated by ChatGPT; the citations didn’t exist, and the resulting court sanctions became one of the most widely circulated cautionary stories of the era.
Start here
Go deeper
Across all these AI eras, the same shape recurs: a confident claim of genuine autonomy, a real and lasting technique the pursuit actually produced, and a quiet walk-back from the original claim as its limits became clear. None of this is a case for cynicism about AI, now or in any earlier decade — it’s closer to the opposite. Every one of these “failures” left something durable behind. The oldest and simplest version of this same idea — a process that does exactly what it’s told and none of what’s meant — is even older than any of these eras: Goethe told a version of it in a 1797 poem, “Der Zauberlehrling” (“The Sorcerer’s Apprentice”), illustrated by artists more than a century before Disney’s famous 1940 adaptation.
The claim in the BDI section above (publication counts rising from the late 1980s, peaking 2013–2017, declining since) is based on original research conducted for this project, not on a pre-existing published source. It’s included here in full so readers can evaluate the method themselves.
Method: the Semantic Scholar Graph API’s bulk search endpoint was queried for papers matching the phrase “BDI agent belief desire intention,” in 5-year windows from 1988 through 2025. A second, related phrasing (“belief desire intention architecture”) was queried as a consistency check. A third, broader query (“multi-agent systems”) was run for comparison.
Results:
| Years | “BDI agent belief desire intention” | “belief desire intention architecture” |
|---|---|---|
| 1988–1992 | 0 | 2 |
| 1993–1997 | 21 | 18 |
| 1998–2002 | 74 | 58 |
| 2003–2007 | 209 | 124 |
| 2008–2012 | 229 | 125 |
| 2013–2017 | 266 | 140 |
| 2018–2022 | 131 | 82 |
| 2023–2025 | 75 | 46 |
The broader query “multi-agent systems” was not found to decline — it climbed monotonically over the same period (338 papers in 1988–1992 to 22,879 in 2023–2025), almost certainly because the broad umbrella term absorbed a large and growing body of unrelated LLM-based multi-agent research beginning around 2023. This is presented as the likely explanation, not a proven causal claim.
Caveats, stated plainly: keyword-matching publication counts are an imperfect proxy for research activity. Some of the apparent decline in BDI-specific terminology may reflect researchers adopting newer terms for closely related ideas, rather than genuine decline in the underlying research area. The consistency of the shape across two different phrasings of the same underlying concept is some evidence the pattern is real rather than an artifact of one query’s exact wording, but this has not been checked against a more rigorous bibliometric database (e.g., Web of Science) and should be treated as a suggestive data point, not a definitive finding.