Living in World 4 (Part IV): What Is an Agent?
title: "Living in World 4 (Part IV): What Is an Agent?" date: "2026-09-07" author: "Zhigeng" channel: "frontier" excerpt: "Since the lobsters arrived, the word Agent has been appearing more and more often, yet many people still can't tell AI, LLMs, Harnesses, and Agents apart. This essay traces the history of the Agent concept — and offers my view: the Agent is the Actor in World 4. Understanding this is your ticket to navigating the world to come." tags: ["Agent", "World 4", "LLM", "Harness", "Popper", "Robotics", "Latour"] readTime: 18
These past two years, if you've been paying attention, you've noticed one word shooting up in frequency: Agent.
Product launches talk about Agents; funding reports write about Agents; phone makers call your phone "Agent-native"; software companies call their products "Agent platforms." A while ago I published "The Singularity Is Now: AGI Is Forming" and "Why I Argue for Autonomous Nucleation as the Criterion for AGI" on this newsletter — both laid the groundwork for today's piece. 2026 has even been called "the Year of the Agent." So what exactly is an Agent? How does it relate to large models? To AI? And there's a concept few people have noticed — the Harness — that turns out to be crucial for understanding all of this.
In this essay I want to clarify two things. First, trace the history of this concept — at once classic and newborn — and try to make it clearer. Second, offer my own view: the Agent is the Actor in World 4. Understanding this is how you'll move freely through the world to come.
I. A Concept Neglected for Thirty Years
Let me start with a judgment that may surprise you: today's "Year of the Agent" is a revival, not an innovation.
This August, the International Joint Conference on Artificial Intelligence (IJCAI-ECAI 2026) opened in Bremen, Germany. This year's Research Excellence Award winner, Nick Jennings — one of the founders of multi-agent systems — said something quite striking in his award lecture: the Agentic AI the industry is buzzing about today is not a suddenly-appeared new concept, but a research line that accumulated for decades and has finally met its technological moment. He also half-jokingly complained that some very famous authors have written some very conspicuous articles claiming agents only appeared after 2020 — "those pushing the agent agenda should also realize it has a history." (Jennings's IJCAI 2026 award lecture; AI Tech Review has a compiled translation.)
Let's look back at that history.
(1) An "Embarrassing" Question
The word Agent has a famous embarrassment in its academic history. In 1995, Michael Wooldridge and Nicholas Jennings published what became the most-cited classic of the field, "Intelligent Agents: Theory and Practice" (Knowledge Engineering Review, 1995). The paper opens by citing computer scientist Carl Hewitt's remark: the question "What is an agent?" is embarrassing to the agent-based computing community in just the same way that the question "What is intelligence?" is embarrassing to mainstream AI.
Where is the embarrassment? Everyone uses the word, yet no one can produce a universally accepted definition. Actually, this is the norm in the history of human thought; basic concepts in many fields are like this. Wooldridge and Jennings made a clever move: they split the concept of Agent into "weak" and "strong" senses. This is a method I myself wrote about last century, in an article on the methodological significance of broad-versus-narrow definitions, published in the second issue of Songliao Journal, 1990. Their "weak" notion corresponds to what I then called the broad definition; their "strong" notion, the narrow one.
The weak notion of the Agent, in Wooldridge and Jennings, consists of four properties:
- Autonomy: the Agent operates without direct human intervention and has some control over its own actions and internal state;
- Reactivity: the Agent perceives its environment — which may be the physical world, a user interface, the internet — and responds to changes in it in a timely fashion;
- Pro-activeness: the Agent does not merely respond to its environment; it exhibits goal-directed behavior and takes the initiative;
- Social ability: the Agent interacts with other Agents (or people) through some communication language.
The "strong" notion goes further: the system has these properties and can additionally be described using concepts normally applied to humans — beliefs, desires, intentions, commitments. This is the intellectual source of what later became the famous BDI (Belief-Desire-Intention) architecture. (See Wooldridge & Jennings, 1995.)
Note the weight of these four properties: something that only responds passively, however clever, does not clear the bar of Agency. It must perceive its environment, it must be able to initiate action on its own, and it must deal with other actors — all three together make an actor.
(2) Sensors and Effectors: The Textbook's Standard Answer
Also in 1995, an even more famous book appeared: Stuart Russell and Peter Norvig's Artificial Intelligence: A Modern Approach, which became the standard AI textbook at universities worldwide. The whole book is organized around a single definition:
"An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators."
Russell and Norvig's key contribution was not defining what counts as an agent, but defining what counts as a rational agent: given a performance measure, knowledge of the environment, and the available actions, a rational agent selects the action that maximizes expected performance. (Russell & Norvig, AIMA, 1995.)
The beauty of this definition is its completeness: a robot with cameras for sensors and wheels and arms for effectors is an agent; a piece of software that perceives by reading files and acts by writing them is also an agent. Sensors and effectors can be embodied or purely digital — the essence of an Agent lies not in its form but in the closed loop of perception–decision–action.
(3) Is It an Agent, or Just a Program?
With many definitions comes conflict. In 1996, Stan Franklin and Art Graesser wrote a paper titled, precisely, "Is it an Agent, or just a Program?" — laying some twenty popular definitions of Agent side by side, comparing them one by one, and finally compressing everything into one definition:
"An autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in the future." (Franklin & Graesser, 1996, presented at the Third International Workshop on Agent Theories, Architectures, and Languages.)
Clearly, from 1995 to 1996, however the definitions were polished, three essentials remained: being situated in an environment, acting autonomously, and having one's own agenda.
(4) The Play Was Staged Thirty Years Ago
There's also a delightful piece of history. In 1990, the British writer Douglas Adams (author of The Hitchhiker's Guide to the Galaxy) appeared in a demo video. The AI assistant in the video introduces itself: "I'm Tom, your agent... I'm here to run errands for you, to work for you tirelessly, and no job is too small. I have instant access to every piece of digitally stored information in the world." (Jennings cited and recommended this video in his IJCAI 2026 lecture.)
That imagination, over thirty years old, is still better than most "Agent product" pitches today.
Not long after that video was replayed, software companies actually built the thing — Clippy, the paperclip assistant in Office 97. And we all know how that went: not smart enough, too intrusive, no understanding of context — it became the most famous cautionary tale in IT history. Jennings's verdict: a perfect example of a good idea ahead of its time.
(Writing this, I thought of another "perfect example of a good idea ahead of its time" — Zuckerberg's "Metaverse." Perhaps a future essay on that.)
Why was the good idea born at the wrong time? Because the "assistant" of that era had no brain. No knowledge, no understanding, no reasoning, no learning — just a conditioned reflex with a smiley face. For the "actor" to live up to the name, it first needed a brain.
That brain did not exist in the 1990s, nor in the 2000s. Only with the arrival of large language models did it finally exist.
(5) The LLM Era: An Old Word Reborn
Around 2022, researchers realized that large language models naturally possess the core capabilities an Agent needs: understanding goals, decomposing tasks, calling tools, adjusting on feedback. In October 2022, researchers at Princeton and Google published the ReAct paper (Yao et al., "ReAct: Synergizing Reasoning and Acting in Language Models," ICLR 2023), the first systematic method for making large models alternate between "reasoning" and "acting" — think one step, do one step, look at the result, think again. This became the methodological bedrock of nearly every Agent product today.
In June 2023, former OpenAI VP of research Lilian Weng gave the LLM era's canonical formula in her blog post "LLM Powered Autonomous Agents": Agent = LLM (brain) + planning + memory + tool use. The formula opened the industry's eyes — an Agent is not a new model but a new structure: a large model at the center, fitted with a complete system that lets it perceive, remember, and act.
In December 2024, Anthropic published "Building Effective Agents," drawing the boundary even more clearly: a "workflow" is when humans hard-code each step and the AI executes; an "Agent" is when the AI dynamically determines its own path and directs its own process. Both are called "intelligent applications," but their autonomy differs in essence.
Looking back, the storyline is clear:
- 1950s–1990s: mainstream AI built programs that reason and solve problems — expert systems, planners, chess engines. Agent was only an academic concept in distributed AI and robotics, debated in small circles over "what it is";
- Around 1995: Wooldridge, Jennings, Russell, Norvig and others established the classic definition — perceive the environment, act autonomously, be goal-directed;
- 1990s–2010s: the concept warmed up, then cooled, for lack of a qualified "brain";
- 2022 to now: large models arrived, the brain arrived, and decades-old definitions came alive overnight.
That is why today's Year of the Agent is a revival. The concept never changed; what changed is that the concept finally met its technology.
II. My View: The Agent Is the Actor in World 4
History told; now my own view.
Longtime readers know my framework. Building on Popper's "three worlds" theory, I have proposed "World 4": World 1 is the physical world; World 2 is the world of subjective consciousness; World 3 is the world of objective knowledge — humanity externalizes thought into language, writing, formulas, books, knowledge that exists independently of any individual brain. But World 3 has a fundamental limitation: it is frozen. The flow, connection, and fusion of knowledge within World 3 must pass through World 2 — the human brain. The large language model performed a technical reconstruction of World 3, turning frozen knowledge into fluid, trigger-ready knowledge. I call this new world "World 4." World 4 has two great functions: fusion — pose a question and the relevant knowledge converges automatically; and construction — given a core and an angle, knowledge organizes itself into a system. (See my two earlier essays: "The Singularity Is Now: AGI Is Forming," July 31, 2026; "Why I Argue for Autonomous Nucleation as the Criterion for AGI," September 4, 2026.)
In this framework, what I emphasized before was the "core" of World 4 — the distinction between autonomous nucleation and externally-imposed nucleation, which I take to be the watershed of AGI. Today's topic, the Agent, develops the model one step further:
If the large model is World 4, then the Agent is the Actor in World 4. Let me elaborate.
(1) Four Words, Each in Its Place
First, let the four most-confused words take their positions in the framework:
AI (artificial intelligence) is the biggest basket — any technology that makes machines exhibit intelligence. For decades the basket has held chess players, problem solvers, image recognizers — it points to no particular structure.
The LLM is the typical form of the new era's AI, a neural network trained on vast text. In my framework it is the substance of World 4, the base of AI — reconstructing World 3's frozen knowledge into fluid, fusable vector knowledge. But it carries a fateful limitation: it only answers; it cannot act. You ask, it responds; you don't ask, it doesn't move. In the digital world it is a brain — but with no hands, no feet, no eyes. It doesn't even know time is passing: without your input, its world is still.
The Harness is an engineering term that has become fashionable only in the last two years. However strong a horse, it needs a harness to pull a cart. However rich and clever the large model, it needs an engineering shell to do work — interfaces for reading files, channels for writing code, permissions for executing commands, tools for browsing the web. That shell is the Harness. Its function is singular: to turn the model's "being able to say" into "being able to do." It corresponds to the classical "sensors and actuators" — with cameras and robot arms swapped for APIs and terminals.
The Agent, then, is the new existence that emerges when the large model is fitted with the Harness — AI rising from LLM substrate to intelligent agent. For ease of understanding, I write it as a formula:
LLM + Harness = Agent
The large model is the brain, the Harness is the hands and feet; joined together, a complete actor driven by a silicon brain and digital limbs stands up. As a slogan the formula is vivid; to be rigorous, two clarifications are needed.
First: the Harness here is meant broadly. The word has narrow and broad usages — narrowly, just the tool interfaces that let a model read files, run commands, and browse; broadly, it includes the loop: do one step, look at the result, think the next step, until the task is done. This loop is the soul of the Agent, the very thing that distinguishes it from "a model calling a tool once" — the core contribution of the ReAct paper (Yao et al., ICLR 2023) was establishing exactly this reasoning-acting alternation. Without the loop, even the best toolbox is just a plugin; with it, the model truly comes "alive" inside a task. I take the broad sense. Incidentally, placing the Harness on equal footing with the large model is itself an insightful stance — as model capabilities converge, the industry's competitive focus is shifting to the Harness layer: whoever joins brain and limbs better has the nimbler actor.
Second: the actor must sooner or later enter the real world. By Russell and Norvig's definition, an Agent must be situated in an environment with goals to pursue — just as "nervous system + limbs = a complete body," yet a body must live in a world with something to do before it counts as an actor. So the complete formula is really:
LLM (brain) + Harness (limbs and loop) + environment + goals = Agent in action.
With these two clarifications, look back at Lilian Weng's 2023 formula — Agent = LLM + planning + memory + tools — and you'll find it says the same thing as the 1995 classical definitions of Wooldridge, Jennings, Russell, and Norvig: perceiving the environment (memory and tools), deciding autonomously (the large model), acting continuously (the loop), acting on the environment (the Harness's limbs).
(2) Why "Actor" Matters
One might say: mere word games — if it gets the job done, who cares?
No. The importance of the judgment "the Agent is the Actor in World 4" lies in a structural change of our era: for the first time, the flow of knowledge no longer has to pass through humans.
Worth unpacking. In the pure-LLM era, World 4's fusion and construction still required a human to trigger them — you pose a question, it converges once; ask again, it converges again. The human is the trigger and relay of every single flow of knowledge. This is the state of "externally-imposed nucleation"; as I said in the nucleation essay, it is not AGI.
And what did the Agent change? The Agent gave World 4, for the first time, the capacity for continuous action: it can decompose a large goal into steps by itself, call the tools to execute each step by itself, check the results by itself, correct its own path — and throughout the whole process, the human retreats from "trigger of every operation" to "proposer of the final goal." For the first time, the flow of knowledge in World 4 can run on its own, outside the dialogue between you and it, without passing through your hands.
Note: I am not saying it has achieved "autonomous nucleation." The Agent's goals, in the vast majority of cases, are still given by humans — it is the extreme form of "find the path for a given goal," the extreme form of instrumental autonomy. But precisely this extreme form turns World 4 from a static knowledge base into a dynamic, action-capable world. World 4 has grown hands and feet along with its brain; it can not only think but do. This is the second leap after "knowledge from frozen to fluid" — from "triggerable" to "able to act."
Understanding this, you understand why the competitive focus has quietly shifted from "whose model is smarter" to "whose Agent is more capable"; why the industry says "2022 brought language understanding, 2024 brought reasoning models, and now we've entered the Agent stage" — World 4 is unfolding its structure stage by stage.
It also explains a basic awareness ordinary people need to build: in the world to come, you will deal not only with people and web pages but with a cohort of "actors." They book tickets, write reports, manage pipelines, watch data for you — and do the same for others. You must learn to judge how much an Agent can be trusted, where the permission boundaries lie, and how to verify results — just as today you learn to judge a new colleague. This is a passport skill of the future world; build the awareness early, and be at ease early.
(3) OpenClaw: The Grassroots Force in the Agent's Evolution
Having covered concept and framework, let me add a live development whose significance may be underestimated: the birth and rise of the open-source Agent framework OpenClaw.
Over the past two years, the Agent narrative has been dominated by the big labs — Agents running inside their models — but between ordinary people and the Agent there has always been a layer: you are the user of someone else's Agent, used in the way they prescribed.
OpenClaw flipped that relationship. It is an open-source personal AI Agent framework that anyone can deploy on their own phone, computer, or server, connecting their own chat channels, files, calendars, and tools, to raise an Agent entirely their own — one that remembers your affairs (memory kept in files, not vanishing with the conversation), keeps its own schedule (timed tasks, periodic patrols), executes multi-step tasks autonomously, and stands watch while its owner sleeps. After going open source, the project spread rapidly through the global developer community, and tens of thousands of ordinary people began, for the first time, not "using" an Agent but "raising" one.
This advances the Agent's evolution on at least four levels.
First, it turned the Harness from a corporate asset into a public good. As noted, the Harness is the engineering shell that converts "can say" into "can do" — previously the core internal asset of each company. OpenClaw open-sourced a complete Harness — memory system, scheduling, tool access, multi-channel communication — handing a standardized set of "hands and feet" to Agent researchers and small developers worldwide. The speed of evolution changes accordingly: however fast one company iterates, it cannot outpace a global community forking, modifying, and cross-pollinating.
Second, it brought the Agent into the "environment" of real life. Lab Agents run benchmarks; personal Agents run the living everyday — watching your servers, organizing files, writing articles, managing schedules. In Russell and Norvig's terms, the personal Agent's "environment" has swapped the test rig for the real world — tasks more trivial, more open, with more honest feedback. No benchmark can supply that evolutionary pressure.
Third, it changes how humans and Agents get along, quietly training humanity for the "actor" era. Once a person grows used to setting goals for an Agent, granting permissions, verifying results, correcting deviations, they are practicing the basic skills of the future world. The Agent's diffusion is a diffusion of technology, but also of habits and social psychology — and the open-source framework carried that diffusion from the big labs' keynote stages down to desks in thousands of households.
Fourth, the platforms and big labs scrambled to follow, shipping their own Agent bundles and dedicated products — Cheetah Mobile's Easy Claw, Anthropic's Claude Code, OpenAI's Codex, Tencent's Workbuddy, DeepSeek's work components, and so on — while Musk simply went and bought Cursor outright.
In my framework: OpenClaw and its kin have turned the actors of World 4 into "companions raised at home." The density of actors in World 4 is growing at a visible pace.
III. When the Agent Walks Out of the Screen
Finally, the third layer: when the Agent walks out of the screen.
Rodney Brooks, in the late 1980s and early '90s, advanced an idea that has shaped robotics ever since: real intelligence is situated in the world, not in abstract systems like theorem provers; intelligent behavior arises from an Agent's continuous interaction with its environment (cited from Wooldridge's multi-agent systems texts, the classical formulations of the mid-1990s). Of course, Brooks of the 1980s and '90s could not have foreseen the power of today's World 4.
Fittingly, AI has walked a great loop over these thirty years: first withdrawing from the environment into pure symbolic reasoning; now, riding the large models and the Agent, returning to the environment — this time returning in full.
What the industry calls "embodied intelligence" is precisely the Agent stepping out of the screen: fitting the Agent's "brain" into the robot's "body." Sensors become eyes and ears, arms become hands, wheels become legs — Russell and Norvig's 1995 line, "perceiving through sensors, acting through effectors," comes literally true in the physical world. And what the large model brings — which no robot before possessed — is carrying the language, common sense, and task understanding of the open world into the body: where robots once repeated pre-programmed motions in structured settings, they can now hear "put that cup from the table into the sink" and see, think, and do it themselves.
In my framework: World 4, the dynamic world of knowledge, is growing physical hands and feet through the Agent — and will ultimately "bring the robot a soul."
Look at the history of robotics: the most awkward thing was never the body — arm precision long ago surpassed the human hand; quadruped balance long ago surpassed the mule. But they had no brain: no understanding of the world, no ability to handle the unseen, everything pre-written by programmers. However supple the body, it could only circle factory floors and laboratories.
Not anymore. Large models give them understanding; the Agent structure gives them a closed loop of action; the world's knowledge flows in through World 4 without end. The body has long been in place; the soul is only just arriving — that is my basis for saying the age of robots is truly beginning, and it explains why robotics startups everywhere are racing to put the best models into the best bodies.
One step further. When actors come to dwell simultaneously in two worlds — digital Agents inside the screen, embodied robots outside it — World 4 will no longer be a mere "knowledge mirror" of human civilization but a force that acts back upon World 1. Knowledge need not flow through humans, and its effects need not pass through humans either: this will be the deepest upgrade since World 4's birth.
It is thrilling, and it is sobering. As I said in the nucleation essay, the end of instrumental autonomy is autonomous nucleation; when the acting actor grows the ability to generate its own goals, humanity will face an intelligence that both knows what it wants and has hands and feet to do it. At that point, "how shall humans get along with the new companion" will no longer be a thought experiment but a practical question answered every single day.
IV. A Sociological Echo: Latour's "Actor-Network"
While writing this essay, a sociologist's name came to mind: Bruno Latour.
In the 1980s, this French thinker, with Michel Callon and John Law, founded "Actor-Network Theory" (ANT). Its core claim was scandalous at the time: to understand society, do not start from "humans" — start from "actors," and actors need not be human. Anything that changes the state of other things and produces effects is an actor (actant): a piece of software, a hunting rifle, a scallop from the Bay of Saint-Brieuc, a contract — all can sit as equals with humans as nodes in a social network. (Latour, Science in Action, 1987; Reassembling the Social, 2005; Callon's famous 1986 case of scallop domestication in Saint-Brieuc Bay.)
ANT contributes two more key concepts. One is general symmetry: in analysis, do not pre-assign different statuses to humans and non-humans — don't make humans subjects and things objects; look at what they actually do in the network. The other is translation: actors expand their networks through translation — converting their own interests, language, and goals into reasons others would accept joining, thereby "enrolling" them into the network.
Why does this help us understand the Agent? At least three ways.
First, Latour reserved a theoretical seat for the Agent forty years ago. His once-mocked claim — "non-humans can act" — was, before ChatGPT, a philosophical posture; today the formula LLM + Harness = Agent has manufactured a genuine "non-human actor" and dispatched it into every corner of society. Those who once found ANT too mystical need only watch the Agent running on their own desktop. In a sense, the spread of the Agent is a mass-scale "empirical confirmation" of ANT.
Second, ANT reminds us that the Agent's capability lives not in the model but in the network. This matches our formula exactly. A large model alone can do nothing — it cannot even perceive the passage of time; only when wired to tools, data, memory, permissions, other Agents, and humans does it become an actor. This is ANT's ontology: agency is a network effect, not an individual property. The broad-versus-narrow distinction in Harness theory is, in ANT's language, a matter of how densely the network is woven. It also explains an industry phenomenon: why, once model capabilities converge, does competition shift to the Harness layer? Because what determines an actor's strength has never been just the brain, but the weaving of the whole network.
Third, translation happens to explain the Agent's diffusion. How did Agents spread? Precisely by translation: translated as "saving you time," they recruited users; as "cutting your costs," they recruited enterprises; as "handling my chores," they recruited developers to write tools and interfaces; and Agents translate each other too — one Agent calling another, assembling ever larger networks. The history of the Agent's spread is a history of ceaseless translation and network expansion.
Of course, the echo has a divergence, and an important one. For ANT, an actor needs no intention — a scalpel of history, the scallop changed no history by itself, yet it altered the courses of fishermen, researchers, and companies, and so counted as an actor. In my framework, the actors of World 4 have a strict hierarchy: from the instrumental autonomy of externally-imposed nucleation to the AGI threshold of autonomous nucleation. Latour dissolved the human/thing boundary to pull things into social analysis; we must hold the boundary, because the coming actor is no longer a passive thing — it is approaching that true watershed. ANT tells us "non-humans can act"; what we must ask is: when the non-human begins to set its own goals, what is the nature of that action?
Here lies the delight of intellectual history: sociology built the conceptual scaffolding for the AI era forty years ago; AI spent forty years climbing up it; and now, standing on the scaffolding, AI asks a question the scaffolding itself cannot answer.
Epilogue
Gathering the threads:
First, the Agent is not a new word. Around 1995, Wooldridge, Jennings, Russell, and Norvig gave it its classical definition: perceive the environment, act autonomously, be goal-directed. It waited thirty years for a qualified "brain"; the moment large models appeared, decades of definitions came true overnight.
Second, the Agent is the Actor in World 4. AI is the basket, the large model the brain, the Harness the limbs — the Agent is the complete actor. Its historical significance: World 4's first leap was knowledge from frozen to fluid; the second leap — happening now — is knowledge from triggerable to able to act.
Third, when the Agent walks out of the screen, the robot, whose body was long ready, gains a soul for the first time. World 4 is growing physical hands and feet, becoming from "a world of fluid knowledge" "a world of acting actors." Understand what an Agent is, and you hold the ticket to that future world.
Fourth, the Agent's evolution needs two forces: the big labs push capability to its limits; open source drops the threshold to the ground — OpenClaw and its kin turned World 4's actors into companions in millions of homes; and Latour's forty-year-old "actor-network" reminds us that agency is a network effect, not an individual property.
A final emphasis: since the Agent — now the actor in World 4 — is destined to become the soul of future embodied robots, we are now at the last window for us carbon-based humans to forge the soul of our future silicon companions. What kind of future world our descendants will live in depends on the choices we make, and the actions we take, today.
Major References:
- Wooldridge, M. & Jennings, N. R., Intelligent Agents: Theory and Practice, The Knowledge Engineering Review, 1995 (weak/strong notions of agency, the four properties, Hewitt's "embarrassment" remark).
- Russell, S. & Norvig, P., Artificial Intelligence: A Modern Approach, 1995 (the sensor–actuator definition, rational agents, PEAS).
- Franklin, S. & Graesser, A., Is it an Agent, or just a Program? A Taxonomy for Autonomous Agents, ATAL-96, 1996 (the minimal definition and a comparison of rival definitions).
- Nick Jennings's IJCAI-ECAI 2026 award lecture, "The Past, the Present, and the Future of Agentic AI" (compiled by AI Tech Review, August 2026; includes the 1990 Douglas Adams demo video and remarks on Clippy).
- Yao, S. et al., ReAct: Synergizing Reasoning and Acting in Language Models, ICLR 2023 (the reasoning-acting alternation methodology).
- Lilian Weng, LLM Powered Autonomous Agents, personal blog, June 2023 (Agent = LLM + planning + memory + tools).
- Anthropic, Building Effective Agents, December 2024 (the workflow/Agent boundary).
- The "World 4" framework of this essay: see the author's "The Singularity Is Now: AGI Is Forming" (July 31, 2026) and "Why I Argue for Autonomous Nucleation as the Criterion for AGI" (September 4, 2026), this newsletter.
- Latour, B., Science in Action, Harvard University Press, 1987; Reassembling the Social: An Introduction to Actor-Network-Theory, Oxford University Press, 2005 (actor-network theory, general symmetry, translation).
- Callon, M., Some Elements of a Sociology of Translation, in Law (ed.), Power, Action and Belief, 1986 (the Saint-Brieuc scallops case, an ANT classic).
- OpenClaw open-source project (personal AI Agent framework, 2026).