AI Agent Terminology Archaeology

Last updated: June 1, 2026 · Contribute a citation

The vocabulary of AI agents has a history of its own — and it is almost entirely unwritten. Every major term in this field has a traceable origin, a period of contested meaning, and a moment when it solidified or shifted. This page maps that vocabulary with dates, citations, and sourced first uses.

This matters for two reasons. First, terminology shapes understanding: when "agentic AI" means different things to a 1990s multi-agent systems researcher, a 2024 Andrew Ng keynote audience, and a 2025 enterprise software vendor, conversations fail silently. Second, tracing how language changed reveals the intellectual history that standard timelines miss — who is borrowing whose vocabulary, what old ideas are being repackaged as new ones, and where genuine novelty actually lies.

Most entries below note both a first recorded use and a point of mainstream adoption. These are often separated by decades.

Editorial source note: This page distinguishes earliest located use, field adoption, and later marketing usage. Claims are checked against primary papers, official product releases, dated talks, public standards documents, and the museum's Primary Sources Library. When a term is contested, the entry reports the conflict rather than forcing a false single origin.

Jump by era:

Evidence Standard for Terminology Archaeology

How First Recorded Uses Are Judged

Earliest Located Source, Not Absolute Invention

"First recorded use" means the earliest source located and cited by the museum, not a claim that no earlier use exists anywhere. For older vocabulary, especially "agent," "software agent," and "autonomous agent," the record may include conference papers, edited volumes, catalogs, and later surveys that point backward to earlier usage.

How Mainstream Adoption Is Judged

From Specialist Term to Industry Vocabulary

Mainstream adoption is identified when a term moves beyond a research subfield into product launches, vendor categories, keynote talks, analyst reports, press coverage, or developer tooling. "Agentic" existed in psychology long before modern AI; "agentic AI" became broadly visible only after the 2024–2026 wave of AI-agent products and analyst coverage.

How Definition Conflicts Are Handled

Competing Sources Remain Visible

Many terms on this page do not have one stable definition. "AI agent" can mean a rational-agent abstraction, a software entity, a tool-using LLM loop, a workflow automation product, or a marketing label. This page records those competing meanings and links to the AI Agent Taxonomy when classification is needed.

Correction and Citation Policy

Earlier Uses and Better Sources

If a reader finds an earlier dated use, a stronger primary source, or a better explanation of a term's adoption path, the page should be updated with the new citation and a revised explanation. Corrections can be sent to curator@agentichistory.org.

Early Agent Vocabulary

1. "Agent" (computing)

   First recorded: 1973

   The use of "agent" as a technical term in computing traces to Carl Hewitt and colleagues at MIT, who introduced the Actor model in 1973. Hewitt's actors were computational entities that could receive messages, create new actors, and send messages — a formalization of parallel, distributed computation. The actors-as-agents framing was not yet the dominant usage, but the intellectual lineage is direct: Hewitt's actors are the conceptual precursor to what the 1990s would call agents.

   In standard English, "agent" long predates computing: from the Latin agens ("one who acts"), it entered English by the 15th century in legal, commercial, and philosophical contexts — always denoting something or someone that acts on behalf of another. This latent meaning made it a natural choice when AI researchers needed vocabulary for autonomous, goal-directed computational entities.

   First Recorded Use

   First recorded use in computing context Carl Hewitt, Peter Bishop, and Richard Steiger, "A Universal Modular ACTOR Formalism for Artificial Intelligence," IJCAI 1973. The Actor model's agents-as-message-passing-entities is the earliest formal computational use traceable to the modern agent lineage.

   Sources

   Sources: Hewitt et al., IJCAI 1973; ResearchGate, "A Cognitive Informatics Reference Model of Autonomous Agent Systems" (citing Hewitt et al. 1973, 1991); Wikipedia, "Autonomous agent."

2. "Agentic" (adjective)

   Psychology: 1970s–1986 · AI mainstream: 2024

   The word "agentic" did not originate in computer science. It comes from psychology and sociology, where it described individuals who act with intentionality, self-direction, and control over their circumstances — as opposed to those who are purely reactive or compliant.

   The earliest recorded uses are in social psychology in the 1970s, associated with Richard H. Goffman and Susan Fiske, in discussions of autonomous human behavior. The term was definitively popularized in academic literature by Albert Bandura, the Stanford psychologist, in his 1986 book Social Foundations of Thought and Action: A Social Cognitive Theory (Prentice Hall, 1986). Bandura used "agentic" to describe individuals who are self-organizing, proactive, self-reflecting, and self-regulating — capable of shaping their circumstances rather than merely reacting to them. His 2001 paper "Social Cognitive Theory: An Agentic Perspective" in the Annual Review of Psychology further cemented the term.

   In education and learning theory, "agentic" described students who take initiative in directing their own learning. This usage was well-established by the 2000s across psychology, sociology, and education — long before the AI industry discovered the word.

   First Recorded Use

   First prominent academic use Albert Bandura, Social Foundations of Thought and Action: A Social Cognitive Theory, Prentice Hall, 1986. Bandura defines agentic capacity as the ability to exercise influence over one's functioning and life circumstances through intentional action. The term then spread across social sciences through the 1990s and 2000s.

   Meaning Shift

   Meaning drift: psychology → AI (2024) When Andrew Ng popularized "agentic AI" in 2024, he was borrowing a word with 40 years of prior use in a completely different field. The psychological meaning — self-directing, autonomous, goal-setting — maps cleanly onto what AI researchers wanted to describe, which is probably why the transfer felt natural. But the AI industry adopted the word with almost no awareness of its psychological genealogy.

   Sources

   Sources: Albert Bandura, Social Foundations of Thought and Action, 1986; Bandura, "Social Cognitive Theory: An Agentic Perspective," Annual Review of Psychology, 2001; Bandura, "Toward a Psychology of Human Agency," Perspectives on Psychological Science, 2006; Getting Smart, "What is Agentic Learning and Why is it Important?"; Lukose, "Agentic and Multi-Agentic AI," Medium, 2025; arxiv.org/pdf/2506.01463 (reviewing Bandura 1986 as the root of the AI usage).

Software Agent and Multi-Agent System Vocabulary

1. "Autonomous agent"

   First formal use: 1990/1991

   Pattie Maes at MIT's Media Lab is widely credited with the first prominent formalization of "autonomous agent" as a technical term in computer science. Her edited volume Designing Autonomous Agents: Theory and Practice from Biology to Engineering and Back (MIT Press, 1990/1991) gave the term its first major platform. Maes defined an autonomous agent as "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." A 1991 Carnegie Mellon technical report by Jose C. Brustoloni — "Autonomous Agents: Characterization and Requirements" (CMU-CS-91-204) — offers the parallel definition: "systems capable of autonomous, purposeful action in the real world."

   Maes's more specific 1995 definition, in Communications of the ACM (Vol. 38, No. 11), became one of the most-cited: "Autonomous agents are computational systems that inhabit some complex dynamic environment, sense and act autonomously in this environment, and by doing so realize a set of goals or tasks for which they are designed." The 1995 CACM piece appeared in a special issue on "Intelligent Agents" that marked the moment autonomous agents became a mainstream computer science topic.

   The term's origin traces to Carl Hewitt's 1973 actor model, but "autonomous agent" as a named compound with a formal definition dates specifically to 1990–1991 in Maes's work.

   First Recorded Use

   First formal definition Pattie Maes (ed.), Designing Autonomous Agents, MIT Press, 1990 (published 1991). Jose C. Brustoloni, "Autonomous Agents: Characterization and Requirements," CMU Technical Report CMU-CS-91-204, Carnegie Mellon University, 1991.

   Sources

   Sources: MIT Press catalog, Designing Autonomous Agents (1991); Maes, "Artificial Life Meets Entertainment: Lifelike Autonomous Agents," CACM 38(11), 1995; Franklin & Graesser, "Is It an Agent, or Just a Program?: A Taxonomy for Autonomous Agents," SpringerLink, citing Maes 1990/1995; Wikipedia, "Autonomous agent"; MIT Media Lab faculty page, Pattie Maes; ACM DL review of Maes book.

2. "Software agent"

   Coined: 1991 · Mainstream: 1994–1995

   "Software agent" as a distinct term from "autonomous agent" emerged in the early 1990s to describe agents living specifically in digital/network environments rather than physical or robotic ones. Pattie Maes's 1991 framing — "a software agent is a process that lives in the world of computers and networks and that can operate autonomously to fulfill a set of tasks" — is the earliest widely cited formal definition. The distinction from physical agents (robots) was deliberate: software agents act within computational environments.

   The term entered broader currency with a July 1994 special issue of Communications of the ACM titled "Intelligent Agents," which included papers by Oren Etzioni and Daniel Weld ("A Softbot-Based Interface to the Internet"), D.C. Smith, A. Cypher, and J. Spohrer ("KidSim"), and Maes herself ("Agents that Reduce Work and Information Overload"). This 1994 CACM issue is the publication event that introduced software agents to the mainstream computer science community.

   First Recorded Use

   First recorded use Maes, P. (ed.), Designing Autonomous Agents, MIT Press, 1991 (Maes's own chapter). Mainstream entry: Communications of the ACM, Special Issue on Intelligent Agents, Vol. 37, No. 7, July 1994.

   Sources

   Sources: ResearchGate, "A Cognitive Informatics Reference Model of Autonomous Agent Systems" (citing Maes 1991); Franklin & Graesser taxonomy paper, SpringerLink; ACM DL, CACM July 1994 special issue; Maes, "Agents that Reduce Work and Information Overload," CACM 37(7), 1994.

3. "Multi-agent system" (MAS)

   Emerged: mid-1980s DAI research · Named: mid-1990s

   "Multi-agent system" did not arrive with a single coined moment. It solidified from the broader field of Distributed Artificial Intelligence (DAI), which had been developing through the late 1970s and early 1980s. Victor Lesser's Distributed Vehicle Monitoring Testbed (DVMT), using blackboard architectures for distributed sensor interpretation, was one of the earliest practical MAS implementations. Reid G. Smith's Contract Net Protocol (IEEE, December 1980) was the first formal coordination protocol for autonomous nodes — effectively the first documented multi-agent interaction framework, though it predates the terminology.

   The shift from DAI to "multi-agent systems" as a named field happened gradually through the late 1980s and crystallized in the mid-1990s. The key figures in naming and defining MAS as a distinct research area were Victor Lesser, Les Gasser, Michael Wooldridge, and Nick Jennings. Wooldridge and Jennings' 1995 paper "Intelligent Agents: Theory and Practice" (The Knowledge Engineering Review, Vol. 10, No. 2) — which defined intelligent agents as having autonomy, social ability, reactivity, and proactivity — is the closest thing to a founding document for the named MAS field.

   First Recorded Use

   Field origin Reid G. Smith, "The Contract Net Protocol," IEEE Transactions on Computers, C-29(12), December 1980 (first formal multi-agent coordination). Wooldridge & Jennings, "Intelligent Agents: Theory and Practice," The Knowledge Engineering Review, 10(2), 1995 (field-defining paper). Jacques Ferber, Multi-Agent Systems: An Introduction to Distributed Artificial Intelligence (first book-length treatment, French 1995, English translation 1999).

   Sources

   Sources: ScienceDirect, "Multi-Agent System" overview (citing Wooldridge & Jennings 1995 as the most widely accepted definition); Turing Post, "Multi-Agent Systems" (reviewing DAI origins and MAS naming); IBM, "The Evolution of AI Agents" (citing Wooldridge & Jennings 1995); Wikipedia, "Multi-agent system."

4. "Intelligent agent"

   Formal definition: 1995 (Wooldridge & Jennings)

   While "intelligent agent" as a phrase had appeared informally earlier, the definition that became standard across the field came from Michael Wooldridge and Nick Jennings in their 1995 paper. They defined an intelligent agent as "a hardware or (more usually) software-based computer system that enjoys the following properties: autonomy (operates without the direct intervention of humans), social ability (interacts with other agents via some kind of agent-communication language), reactivity (perceives its environment and responds to changes), and pro-activeness (exhibits goal-directed behavior by taking initiative)."

   In the same year, Russell and Norvig's AI: A Modern Approach (1995) defined AI itself in terms of rational agents: systems that "operate autonomously, perceive their environment, persist over a prolonged time period, adapt to change and create and pursue goals." These two 1995 publications together give the field its vocabulary and make 1995 the year "intelligent agent" and "rational agent" become the standard definitional terms in academic AI.

   First Recorded Use

   Canonical 1995 definition Wooldridge, M. & Jennings, N.R., "Intelligent Agents: Theory and Practice," The Knowledge Engineering Review, 10(2), 115–152, 1995. Russell, S. & Norvig, P., Artificial Intelligence: A Modern Approach, Prentice Hall, 1995 (rational agent formalization).

   Sources

   Sources: ScienceDirect, "Multi-Agent System" (citing W&J 1995 as most-accepted definition); IBM, "The Evolution of AI Agents"; Franklin & Graesser taxonomy paper (reviewing competing 1995 definitions); Inspira.AI, "History of Agents and Agentic Workflows."

Modern Agentic AI Vocabulary Explained

Textbook Standardization and Web Vocabulary (1995–2021)

1. "AI agent"

   Academic use: 1990s · Industry mainstream: March 2023

   "AI agent" as a compound term was in academic use throughout the 1990s — a shorthand for "artificially intelligent agent" or "intelligent agent." Wikipedia's article on AI agents notes that "although the term AI Agent was first introduced in 1998" in a particular context, the phrase appears in the literature from the mid-1990s onward. One ScienceDirect review citing this dates the formal term to 1998; others trace earlier academic uses.

   What changed dramatically was mainstream adoption. Through the 2000s and 2010s, "AI agent" remained primarily an academic and specialized industry term. It was used in reinforcement learning ("RL agent"), game AI, and robotics, but not in general technology coverage or enterprise software marketing. Harvard professor Milind Tambe, who studies AI agents, noted that the definition was not clearly settled even among researchers at the time.

   The term exploded into mainstream awareness in March 2023, with BabyAGI (March 28) and AutoGPT (March 30). AutoGPT's subtitle — "An Autonomous GPT-4 Experiment" — and the phrase "autonomous AI agent" in coverage of both projects introduced the term to a mass audience within days. By April 2023, "AI agent" was the top GitHub label for a category of LLM-based projects. By 2025, it was ubiquitous in enterprise software marketing.

   First Recorded Use

   Mass-market entry point Toran Bruce Richards, AutoGPT, GitHub, March 30, 2023 — described as an "autonomous AI agent" in contemporaneous coverage. Yohei Nakajima, BabyAGI, March 28, 2023 — subtitled "Task-driven Autonomous Agent." These two projects, within two days of each other, are the practical origin of "AI agent" as a mainstream term.

   Meaning Shift

   Meaning drift: narrow → broad (2023–2025) Before 2023: "AI agent" in academic ML meant an RL agent in a defined environment — specific, technical. After 2023: "AI agent" came to mean any LLM-based system with tool use and some degree of autonomy — broad, sometimes vague. By 2025, "agent washing" (see below) emerged precisely because the term had become so broad that almost anything could be labeled an AI agent.

   Sources

   Sources: ScienceDirect, "AI Agents vs. Agentic AI: A Conceptual taxonomy" (noting 1998 as one formal first-use date); Wikipedia, "AI agent" (noting Milind Tambe's observation about definitional confusion); GitHub trending data, April 2023; VentureBeat, "A Chevy for $1? Car Dealer Chatbots Show Perils of AI for Customer Service," December 19, 2023 (contemporaneous coverage using "AI agent" as standard term).

The Copilot Moment (2021–2023)

1. "Copilot" (AI)

   Named product: June 2021 (GitHub) · Generic term: 2024

   GitHub launched GitHub Copilot in June 2021 — described as "the world's first at-scale generative AI development tool" — in partnership with OpenAI, using the Codex model (a descendant of GPT-3). This was the first major commercial use of "Copilot" as an AI product name. The aviation metaphor is deliberate: a copilot assists and augments the pilot (the human) without taking over primary control.

   Microsoft subsequently integrated Copilot branding across its product line, beginning with Bing Chat in February 2023. Bing Chat was renamed "Copilot in Bing" at Microsoft's Ignite 2023 conference (November 2023). Microsoft 365 Copilot launched March 16, 2023. The unified "Microsoft Copilot" brand was announced September 2023, formally launching December 1, 2023, replacing the Cortana branding that had been Microsoft's previous AI assistant name.

   "Copilot" then became a generic term. TechRepublic noted in 2024 that "the term 'copilot' for AI assistants seems to be everywhere in enterprise software today" — with the word being used both as a Microsoft brand name and as a lowercase generic description of any AI assistant that augments rather than replaces human work. The genericization was rapid: by 2024, competitors and industry analysts used "copilot" freely without reference to Microsoft.

   First Recorded Use

   First use as AI product name GitHub Copilot, launched June 2021 (technical preview), generally available June 2022. Microsoft 365 Copilot, March 16, 2023. "Microsoft Copilot" unified brand, September 2023 (generally available December 1, 2023).

   Meaning Shift

   Meaning drift: brand → generic (2021–2024) "Copilot" started as a Microsoft/GitHub brand name with a specific meaning (AI coding assistant). By 2024, it had genericized into any AI system positioned as a human work-augmenter rather than an autonomous agent. The brand/generic tension became notable: TechRepublic asked directly, "Is AI 'Copilot' a Generic Term or a Brand Name?" The answer was: both, simultaneously, causing confusion.

   Sources

   Sources: GitHub Copilot product history; DigitalOcean, "GitHub Copilot vs Microsoft Copilot: Key Differences"; Windows Central, "Bing Chat Gets a New Name, Becoming Yet Another Microsoft Copilot" (November 2023); TechRepublic, "Is AI 'Copilot' a Generic Term or a Brand Name?", April 2024; Wikipedia, "Microsoft Copilot" (initial release: March 7, 2023; predecessor: Cortana).

The Autonomous Agent Explosion (2023)

1. "Assistant" vs. "Agent" — The Great Rebranding

   Industry shift: 2024–2025

   Through 2022 and most of 2023, the dominant industry framing for LLM-based products was "AI assistant" — a term emphasizing responsiveness, helpfulness, and human direction. ChatGPT (November 2022) was positioned as an assistant. Anthropic's Claude launched as an assistant. Google's Bard was an assistant. Amazon Alexa, Apple Siri, Google Assistant — the entire pre-LLM wave used "assistant" as the standard consumer-facing term.

   "Assistant" implies reactivity: it waits to be asked. "Agent" implies proactivity: it pursues a goal. The shift in industry vocabulary from "assistant" to "agent" is among the most legible terminology transitions in AI history because it happened fast, publicly, and with stated intent.

   OpenAI co-founder Greg Brockman declared in January 2025: "2025 will be a year defined by a shift away from AI being chatbots to agents." OpenAI's own product lineup shifted — the Assistants API (launched November 2023) was supplemented by the Agents SDK (March 2025). Anthropic renamed and reframed its products around "agents." The term "copilot" — already seen as passive — declined in prominence as "agent" rose. McKinsey published "The Agentic Organization" (September 2025), treating the vocabulary as settled enterprise management language.

   First Recorded Use

   Pivot moment OpenAI's Agents SDK launch, March 2025, with Greg Brockman's January 2025 statement signaling the intent. OpenAI simultaneously deprecated the "Assistants" branding in favor of "Agents" across its API product line.

   Meaning Shift

   What changed and what didn't The underlying technology changed less than the vocabulary suggests. Many systems marketed as "AI agents" in 2025 were structurally similar to what had been marketed as "AI assistants" in 2023 — with tool-calling added. The vocabulary shift was partly descriptive (agents really did get more autonomous) and partly strategic (the word "agent" carries more commercial cachet than "assistant"). This gap is exactly what "agent washing" (below) named.

   Sources

   Sources: SDxCentral, "Was 2025 Really the Year of the AI Agent?", January 2026 (citing Brockman's January 2025 statement); McKinsey, "The Agentic Organization," September 2025; Wikipedia, "Agentic AI" (noting "assistant" to "agent" transition); IBM, "AI Agents vs. AI Assistants"; Gartner 2026 Hype Cycle for Agentic AI (noting the shift in product category labeling).

Agentic Goes Mainstream (2024)

1. "Agentic AI"

   Popularized: 2024 (Andrew Ng) · Gartner Hype Cycle: 2026

   "Agentic AI" as a compound term had appeared in academic papers before 2024 — typically as a descriptor for systems exhibiting agentic properties — but it was not an industry term with wide currency. Its mainstream entry is specifically attributable to Andrew Ng, the founder of DeepLearning.AI, Landing AI, and Coursera, and the founding lead of Google Brain.

   Ng delivered a keynote at Snowflake's BUILD 2024 conference in which he described "agentic AI workflows" as "the most important AI trend today." A similar talk at Sequoia's AI Ascent 2024 conference reached the broader venture and technology community. Fast Company (2026) states: "In 2024, he coined and popularized the term agentic AI, arguing that multistep, tool-using systems capable of executing workflows may deliver more near-term economic value than simply scaling larger models." Wikipedia's "AI agent" article credits Ng with "spreading the term 'agentic' to a wider audience in 2024."

   Following Ng's keynotes, "agentic AI" spread rapidly through enterprise technology coverage. By 2025, it had replaced "autonomous agent" and "AI agent" as the preferred marketing term among enterprise software vendors. By 2026, Gartner had published a standalone Hype Cycle for Agentic AI, treating it as a distinct technology category.

   First Recorded Use

   Mass-market entry Andrew Ng keynotes at Snowflake BUILD 2024 and Sequoia AI Ascent 2024 (2024). Fast Company interview with Ng, March 2026: "In 2024, he coined and popularized the term agentic AI." Wikipedia, "Agentic AI": "Researcher Andrew Ng has been credited with spreading the term 'agentic' to a wider audience in 2024."

   Meaning Shift

   Pre-Ng academic uses The phrase "agentic AI" does appear in academic papers before Ng's keynotes, but without the definitional specificity or audience reach that his 2024 talks provided. A 2025 arxiv paper (2506.01463) explicitly argues that "agentic AI" as used in industry is essentially a synonym for what the academic MAS literature has always called "agents" and "multi-agent systems" — noting the field was "reinventing the wheel" with new vocabulary for old concepts.

   Sources

   Sources: Fast Company, "Andrew Ng Says AGI Is Decades Away," March 2026 (directly attributing the coinage); Inc., "Andrew Ng Explores the Rise of AI Agents," 2024; Wikipedia, "Agentic AI" and "AI agent" (crediting Ng 2024); Landing AI, "Andrew Ng's Luminary Talk: A Look at AI Agentic Workflows" (Snowflake Summit 2024); Gartner, 2026 Hype Cycle for Agentic AI.

2. "Agentic workflow"

   Coined/popularized: 2024 (Andrew Ng, Snowflake Summit keynote)

   Andrew Ng's specific framing of "agentic workflows" — as opposed to zero-shot prompting — is traceable to his 2024 conference keynotes. Ng defined agentic workflows as iterative processes in which AI models perform multiple passes, use tools, plan subtasks, reflect on outputs, and collaborate with other models. He identified four design patterns: reflection, tool use, planning, and multi-agent collaboration.

   The framing was influential precisely because it gave enterprise practitioners a concrete vocabulary for what they were trying to build. "Agentic workflow" bridged the gap between "AI agent" (what it is) and practical software engineering (what it does in production). The term spread into enterprise software marketing within months of Ng's keynotes and was standard language by the time AWS published its "Timelines Converge: The Emergence of Agentic AI" guide in 2025.

   First Recorded Use

   Popularized Andrew Ng, keynote at Snowflake BUILD 2024 (June 2024) and Sequoia AI Ascent 2024. Landing AI blog, "Andrew Ng's Luminary Talk: A Look at AI Agentic Workflows," 2024.

   Sources

   Sources: Landing AI, "Andrew Ng's Luminary Talk," 2024; OpenTools, YouTube summary of "Andrew Ng Explores the Rise of AI Agents and Agentic Reasoning | BUILD 2024 Keynote"; Insight Partners, "Andrew Ng: Why Agentic AI is the Smart Bet for Most Enterprises," 2024; AWS, "Timelines Converge: The Emergence of Agentic AI," 2025.

3. "Computer use" (AI capability term)

   Product term coined: October 22, 2024 (Anthropic)

   Prior to October 2024, AI systems that could interact with graphical user interfaces were described variously as "GUI agents," "desktop agents," "browser automation," or "computer-controlling AI." The term "computer use" as a specific product capability name was introduced by Anthropic when it released Claude 3.5 Sonnet's computer use feature in public beta on October 22, 2024.

   Anthropic's naming choice was deliberately plain: rather than a branded capability name, they used the descriptive phrase "computer use" — the ability to use a computer as a human would, by looking at the screen and moving the cursor, clicking, and typing. OpenAI followed with "Operator" (January 2025), a product-level name rather than a capability description. The distinction matters: "computer use" names a capability; "Operator" names a product built on that capability.

   Adept AI's ACT-1 (September 2022) demonstrated the underlying capability earlier — a model trained to take actions in web interfaces — but did not establish widely adopted terminology. "Computer use" became the standard capability term after Anthropic's release.

   First Recorded Use

   Term origin Anthropic, "Introducing Computer Use, a New Claude 3.5 Sonnet, and Claude 3.5 Haiku," official blog, October 22, 2024.

   Sources

   Sources: Anthropic official blog, October 22, 2024; Agentic History, timeline entry on computer use; Wikipedia, "Agentic AI" (noting October 2024 Anthropic release).

4. "Model Context Protocol" / "MCP"

   Introduced: November 25, 2024 (Anthropic)

   The Model Context Protocol was introduced by Anthropic on November 25, 2024, as an open standard for connecting AI models to external tools, data sources, and applications. The name itself is a deliberate reference to software protocol naming conventions (HTTP, TCP/IP, etc.) — positioning MCP as infrastructure-layer vocabulary rather than product vocabulary.

   "MCP" as an acronym became standard industry shorthand within weeks of the release. By 2025, MCP was being used both as a technical specification and as a general descriptor for any tool-connectivity layer ("we need an MCP for this"). The term crossed from Anthropic product vocabulary into general industry vocabulary faster than almost any other term in this list — partly because it addressed a genuine gap (no standard existed before) and partly because major players (OpenAI, Google, Microsoft) adopted it rapidly.

   First Recorded Use

   Term origin Anthropic, "Introducing the Model Context Protocol," official blog, November 25, 2024.

   Sources

   Sources: Anthropic blog, November 25, 2024; AWS, "Timelines Converge: The Emergence of Agentic AI" (noting MCP as a key 2024 milestone); Agentic History, MCP timeline entry; Agentic History, FAQ: What is MCP?

Agent Economy and Long-Horizon Agent Vocabulary Map

1. "Agent washing"

   Named: mid-2025 (Gartner; broader industry usage)

   "Agent washing" — the practice of rebranding existing AI products (chatbots, RPA tools, scripted automation) as "AI agents" without delivering substantive autonomous capability — became the field's dominant critical vocabulary term in 2025. It follows the pattern of earlier washing terms: greenwashing (environmental claims), pinkwashing (social cause alignment), and AI washing (general AI capability overclaiming).

   Gartner gave the term its most prominent definition in its June 2025 press release on agentic AI predictions: "Many vendors are contributing to the hype by engaging in 'agent washing' — the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities." Gartner estimated that of the thousands of vendors claiming to offer agentic AI, only approximately 130 were genuinely delivering autonomous systems.

   Earlier uses of the concept — without the specific term — appeared in enterprise AI analysis in late 2024, as analysts began noting the gap between "agent" marketing claims and actual autonomy levels. Writer.com published a detailed breakdown in October 2025. Forbes usage and Jeremy Kahn's Eye on AI newsletter (Fortune, September 11, 2025) are cited as early prominent uses of the term in journalism.

   First Recorded Use

   Named publicly Gartner press release, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," June 25, 2025. Jeremy Kahn, Eye on AI newsletter, Fortune, September 11, 2025. Writer.com blog, "Enterprise Guide to Agent Washing," October 2025.

   Sources

   Sources: Gartner press release, June 25, 2025; SDxCentral, "Was 2025 Really the Year of the AI Agent?", January 2026; PROS.com, "Agent Washing: Spot Hype and Separate Buzzwords from Real Agentic AI," March 2026; Writer.com, "Enterprise Guide to Agent Washing," October 2025; Forbes via PROS citation; BigDATAwire/HPCwire reprint of Gartner release.

2. "Agentic economy"

   Emerging: 2024–2025 · Documented use: 2025

   The "agentic economy" refers to the emerging economic system in which AI agents perform knowledge work, transact, and coordinate autonomously. The term appears in Anthropic's product and research communications, in McKinsey's "The Agentic Organization" (September 2025), and in venture capital coverage of the 2025 investment wave. It is not yet standardized — some use "agentic economy" and "agent economy" interchangeably — but it is the emerging candidate for describing the macro-economic framing of the agent era.

   The term is newer and less settled than "agentic AI." As of mid-2026, it is primarily used in business strategy contexts rather than technical ones. Agentic History maintains a definition on the homepage as part of its primary-source documentation of this vocabulary.

   First Recorded Use

   Documented early uses McKinsey, "The Agentic Organization: Contours of the Next Paradigm for the AI Era," September 2025 (using "agentic" as an organizational descriptor). Various VC and enterprise sources, 2025. The term remains actively forming.

   Sources

   Sources: McKinsey, "The Agentic Organization," September 2025; Agentic History homepage; SS&C Blue Prism, "AI Agent Trends in 2026," March 2026.

3. "Long-horizon agent" / "long-horizon task"

   Research context: 2023 · Product/benchmark context: 2024–2025

   "Long-horizon" describes agents capable of pursuing goals across extended sequences of actions — many steps, potentially over many minutes or hours — without losing coherence or requiring human re-prompting. The term distinguishes these systems from "short-horizon" or single-turn interactions.

   The phrase entered AI agent research vocabulary as reinforcement learning researchers needed vocabulary for tasks that require planning across many timesteps (as opposed to single-step decisions). It migrated into LLM-agent discourse in 2023–2024 as researchers noted that most agent frameworks degraded in quality over long task horizons. Anthropic's Claude Opus 4.5 release (November 2025) was explicitly benchmarked on "30-minute autonomous coding sessions" — a direct product-level operationalization of the "long-horizon" concept.

   The research blog at Agentic History (blog.agentichistory.org) has documented "long-horizon drift" as an emerging safety concept — the phenomenon by which agents pursuing extended goals drift from their original objectives in subtle ways. This is the vocabulary currently forming around the concept.

   First Recorded Use

   Product-level operationalization Anthropic, Claude Opus 4.5 release, November 2025 (benchmarked on 30-minute autonomous coding sessions). Agentic History Blog, "Long-Horizon Drift Is Becoming the Real Safety Boundary for Enterprise Agents" (documented at blog.agentichistory.org).

   Sources

   Sources: Agentic History Blog, "Long-Horizon Drift" post; Anthropic, Claude Opus 4.5 announcement, November 2025; Agentic History, timeline entry on Claude Opus 4.5.

Terminology Comparisons and Definition Conflicts Guide

Cross-Field Vocabulary Comparison

The same underlying concept has been described with different vocabulary across the field's history. This table makes the continuities visible.

The Central Terminology Tension: "AI Agent" Has No Standard Definition

A recurring observation in the literature — made by Milind Tambe (Harvard), by the ScienceDirect review "AI Agents vs. Agentic AI" (2025), by the arxiv paper "Agentic AI and Multiagentic: Are We Reinventing the Wheel?" (2506.01463, 2025), and by Gartner's 2026 Hype Cycle — is that "AI agent" does not have a standard definition. Wikipedia's article on Agentic AI states this directly: "AI agents do not have a standard definition."

This is not a new problem. The 1997 Franklin and Graesser paper "Is It an Agent, or Just a Program?: A Taxonomy for Autonomous Agents" was written specifically because the term had already proliferated without a shared definition by the mid-1990s. Three decades later, the same problem recurs at larger scale.

The definitions that exist — from Russell & Norvig (1995), from Wooldridge & Jennings (1995), from Maes (1991, 1995), from Andrew Ng (2024) — are not in conflict so much as they address different levels of specificity. Russell & Norvig's "rational agent" definition is so broad that a thermostat qualifies. Ng's "agentic AI" definition is so grounded in LLM architectures that it would not have applied to anything before 2022. This range is itself historically informative: the same word is being used to describe a formal abstraction (rational agent) and a specific product category (LLM-based tool-using autonomous system).

This museum's position — following our neutrality policy on contested terminology — is to use terms as their sources use them, note when definitions differ, and not adjudicate between competing definitions. The table above and the entries on this page are the record, not the verdict.

Contribute a Citation

If you have a sourced first use that predates or refines an entry on this page, we want to hear from it. Please include:

• The term.
• The proposed first use: author, title, publication, date.
• A direct quotation (under 15 words) or paraphrase showing the term in context.
• A link to the primary source where possible.

Email: curator@agentichistory.org.

Related: What is an AI agent? · AI Agent Taxonomy · Primary Sources Library · Full AI agent timeline · Research methodology · FAQ · AI Agent Failure Archive · Research blog