Question 1

What is the history of AI agents?

Accepted Answer

The history of AI agents spans roughly seven decades. The conceptual seed is planted at the 1956 Dartmouth Summer Research Project on Artificial Intelligence. Multi-agent coordination is formalized by Reid G. Smith's Contract Net Protocol in December 1980. Marvin Minsky's Society of Mind (1986) argues intelligence emerges from the interaction of many small agents. Michael Bratman's BDI framework (1987) supplies the dominant agent architecture for the next two decades. Stuart Russell and Peter Norvig's textbook (1995) reorganizes the field around the rational-agent abstraction. The 1990s formalize multi-agent systems in academic terms: KQML (1994), FIPA-ACL (1997), Yoav Shoham's Agent-Oriented Programming, and Dr. Anand Rao's AgentSpeak(L) (1996) and the Jason interpreter. This decade produces an enormous literature that today's LLM-agent frameworks largely re-derive independently. The modern LLM-driven era begins with Adept's ACT-1 demo (September 2022), accelerates with the ReAct paper (October 2022), and explodes with the OpenAI APIs (March 2023) and then BabyAGI and AutoGPT later that month. Computer-use agents from Anthropic (October 2024) and OpenAI Operator (January 2025) mark the next major shift. By 2025, every major frontier lab offers a first-party agent SDK, and investment in AI-agent startups reaches $8.7B in Q1 2025 alone. The full chronology is on the timeline.

Question 2

Who invented AI agents?

Accepted Answer

No single person invented AI agents. The credit is distributed across decades and disciplines, and different people are the correct answer depending on which layer of the technology you mean. The multi-agent coordination model was first formalized by Reid G. Smith in 1980. The theoretical foundation of intelligence as a collective of agents comes from Marvin Minsky in 1986. The BDI architecture — the dominant technical framework for agent software through the 1990s and 2000s — derives from philosopher Michael Bratman's work in 1987. The educational standard that defined AI itself in agent terms was set by Stuart Russell and Peter Norvig in their 1995 textbook. The modern LLM-agent paradigm was defined by the ReAct paper (Shunyu Yao and co-authors, October 2022). The first publicly available autonomous LLM agents were BabyAGI (Yohei Nakajima, March 28, 2023) and AutoGPT (Toran Bruce Richards, March 30, 2023). The 1956 Dartmouth attendees — John McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester — established the broader field that eventually defined itself in agent terms, making them the most distal precursors.

Question 3

Who pioneered AI agents?

Accepted Answer

The pioneers of AI agents fall into two clear groups. Foundational pioneers (1980–1995): Reid G. Smith — Contract Net Protocol (IEEE Transactions on Computers, December 1980). The first formal multi-agent coordination framework. Marvin Minsky — The Society of Mind (Simon & Schuster, 1986). Intelligence as a collective of many agents. Michael Bratman — Intention, Plans, and Practical Reason (Harvard University Press, 1987). The philosophical foundation of BDI agents. Yoav Shoham — Agent-Oriented Programming (Artificial Intelligence, 1993). Agents as the primary unit of software design. Dr. Anand Rao — AgentSpeak(L) (MAAMAW-96, 1996). Operationalizing BDI in an executable programming language. Stuart Russell & Peter Norvig — AI: A Modern Approach (1995). The textbook that cemented the rational-agent abstraction across all of AI. Modern LLM-agent pioneers (2022–2024): Shunyu Yao and the ReAct co-authors — the reasoning-plus-action paradigm (October 2022) that nearly every modern agent framework implements. David Luan and the Adept team (ACT-1) — the first widely publicized transformer trained to take actions in software interfaces (September 2022). Yohei Nakajima (BabyAGI) — first viral autonomous LLM agent (March 28, 2023). Toran Bruce Richards (AutoGPT) — the project that introduced "autonomous AI agent" to the mainstream (March 30, 2023). Scott Wu, Steven Hao & Walden Yan (Cognition) — Devin, the first AI software engineer agent (March 2024). Bret Taylor & Clay Bavor (Sierra) — first enterprise-grade customer-experience agents at scale (2023–2025).

Question 4

Who are the pioneers of multi-agent systems?

Accepted Answer

The canonical pioneers of multi-agent systems are: Reid G. Smith (1980) — the Contract Net Protocol gave the field its first formal coordination framework: manager–contractor bidding over task allocation among autonomous nodes. Marvin Minsky (1986) — Society of Mind gave the field its theoretical foundation: intelligence as a collective. This book is the most widely cited theoretical precursor to modern multi-agent LLM architectures. Yoav Shoham (1993) — Agent-Oriented Programming proposed agents as the primary unit of software design, providing the multi-agent systems field with a software-engineering formalization. Tim Finin and the KQML community (1994) — Knowledge Query and Manipulation Language became the first widely deployed agent communication language. FIPA (from 1997) — the Foundation for Intelligent Physical Agents produced the FIPA-ACL standards, the most widely adopted agent communication specifications before the LLM era. Dr. Anand Rao (1996) — AgentSpeak(L) operationalized the BDI model in a logic-based executable language, bridging theory and practice. The Jason interpreter (Hübner & Bordini) later extended AgentSpeak into a full multi-agent development platform. See the 1980, 1986, and 1990s sections of the timeline for primary citations.

Question 5

Who are the pioneers of modern multi-agent systems?

Accepted Answer

"Modern multi-agent systems" generally refers to LLM-driven multi-agent architectures — the systems that dominate the field from 2022 onward. The pioneers in this narrower sense are: Shunyu Yao and ReAct co-authors (October 2022) — the reasoning-plus-action loop that nearly all modern agent frameworks implement. Without ReAct, there is no coherent template for how an LLM-based agent plans and acts. Harrison Chase and the LangChain team (from late 2022) — the Python and JavaScript framework that productized the ReAct pattern and made LLM-agent development accessible to the broad developer community. Microsoft Research's AutoGen team (2023) — one of the first widely adopted frameworks specifically designed for multi-agent LLM architectures, where multiple models play distinct roles and communicate. The CrewAI team (2024) — popularized role-based multi-agent architectures with structured handoffs between agents playing assigned roles (researcher, writer, critic, etc.). The LangGraph team at LangChain (2024) — introduced graph-based multi-agent orchestration, letting developers define explicit state machines governing agent interactions. Anthropic (November 2024) — the Model Context Protocol gave the agent ecosystem its first serious candidate for a universal tool-connectivity standard, enabling multi-agent systems to compose tools from any MCP-compatible provider.

Question 6

When did AI agents start?

Accepted Answer

Five plausible "start" dates, depending on what you mean: 1956 — the conceptual founding of AI at Dartmouth. The Logic Theorist demonstrates goal-directed machine reasoning. December 1980 — the first formal multi-agent system: Smith's Contract Net Protocol gives autonomous nodes a formal coordination mechanism. 1995 — "agent" becomes the standard definition of AI in mainstream education, via Russell & Norvig. October 2022 — the ReAct paper defines the modern LLM-agent paradigm: reason, act, observe, repeat. March 2023 — autonomous LLM agents enter the mainstream with BabyAGI (March 28) and AutoGPT (March 30), and the ChatGPT API makes agent development economically viable. Most contemporary use of "AI agents" — meaning autonomous, tool-using LLM systems — points at March 2023 as the practical starting point for what the industry now calls the agent era.

Question 7

When did agentic AI begin?

Accepted Answer

"Agentic AI" as a term enters mainstream usage during 2024 and becomes ubiquitous in 2025, as model labs shift marketing from "assistants" to "agents." The capability it describes — autonomous, planning, tool-using LLM systems — begins earlier: with the ReAct paper (October 2022) as the technical origin and the AutoGPT/BabyAGI moment (March 2023) as the cultural origin. By the time the phrase is in common use, the underlying systems have been in development for two to three years. The deeper capability — autonomous agents that plan and act — is decades older. The term "agentic AI" is new; the concept is not.

Question 8

What was the first AI agent?

Accepted Answer

The phrase has many defensible answers depending on definition: First goal-directed symbolic AI program: the Logic Theorist by Newell, Simon, and Shaw (1956). First formal multi-agent coordination system: the Contract Net distributed sensing system (Smith, 1980). First widely demonstrated transformer-based action agent: Adept's ACT-1 (September 2022). First publicly demonstrated end-to-end autonomous LLM agent: BabyAGI (March 28, 2023) and AutoGPT (March 30, 2023), within two days of each other. The question of "firsts" in this field is genuinely contested. See our Neutrality on "Firsts" policy.

Question 9

What was the first autonomous AI agent?

Accepted Answer

The first publicly available autonomous LLM agent that demonstrated the now-canonical loop of objective → planning → execution → reprioritization is BabyAGI, posted by Yohei Nakajima on March 28, 2023. AutoGPT followed two days later (March 30, 2023) and overtook BabyAGI in mindshare almost immediately because it shipped with web browsing, file operations, and code execution out of the box. We treat them as effectively simultaneous origins of the modern autonomous-agent movement. Both were preceded by the ReAct paper (October 2022) that defined their architectural pattern, and by Adept's ACT-1 demo (September 2022) that demonstrated action-taking transformers at a lab level. The deeper roots of autonomous agents run to the Contract Net Protocol (Smith, 1980) and BDI architectures (Bratman, 1987) — formal systems for autonomous goal-directed behavior developed decades before LLMs existed.

Question 10

What was BabyAGI?

Accepted Answer

BabyAGI was an autonomous AI agent script posted to GitHub by Yohei Nakajima on March 28, 2023. The full title of the accompanying paper was Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications. In approximately 140 lines of Python, BabyAGI demonstrated the canonical autonomous-agent loop: a user provides an objective; GPT-4 generates a task list to achieve it; each task is executed by GPT-4 calling LangChain tools; the results are stored in a Pinecone vector database; GPT-4 then reprioritizes the task list based on those results; the loop repeats indefinitely. This objective → task creation → execution → reprioritization → repeat pattern is the same core structure used by AutoGPT, early LangChain agents, and most subsequent open-source autonomous agents. BabyAGI was the first viral, publicly available implementation of an autonomous LLM agent. It was followed two days later by AutoGPT, which shipped with more out-of-the-box capabilities and overtook BabyAGI in public mindshare. Both are cited in dozens of academic papers as the practical origin of the modern open-source autonomous-agent movement.

Question 11

What is the history of agentic AI?

Accepted Answer

The history of agentic AI is the history of AI agents seen through the lens of contemporary marketing. The capability arc runs from Contract Net (1980) to the present. The terminology arc is much shorter: "agentic AI" was a niche academic phrase before 2023 and became the dominant industry framing during 2024 and 2025, as model labs and venture capitalists rebranded "AI assistants" into "AI agents." The capability milestones that define the agentic AI story: Contract Net Protocol (1980), BDI architectures (1987), multi-agent systems literature (1990s), deep RL agents like AlphaGo (2016), GPT-3 API (2020), Adept ACT-1 (2022), ReAct paper (2022), BabyAGI and AutoGPT (March 2023), LangChain and framework explosion (2023), Devin — first AI software engineer (March 2024), Anthropic computer use (October 2024), MCP standard (November 2024), OpenAI Operator (January 2025), Agent SDKs from all major labs (2025). See the full timeline.

Question 12

What is the difference between agentic AI and an AI agent?

Accepted Answer

An AI agent is a specific system: a piece of software that perceives, decides, and acts. Agentic AI is the broader category — the architectural and product approach in which AI systems are built to be agents rather than single-turn chatbots. In practice the terms are used almost interchangeably in 2025–2026. "Agentic AI" usually refers to the field or paradigm — the idea that AI systems should plan, use tools, and act autonomously. "AI agent" usually refers to a specific instance — Devin is an AI agent; the architectural approach that enables Devin is agentic AI. The distinction matters more in academic usage than in industry usage. In Russell & Norvig's formulation, any system that perceives and acts is an agent — which technically includes a chess engine. In 2025 industry usage, "AI agent" and "agentic AI" both almost exclusively mean LLM-driven systems with planning loops, tool use, and memory.

Question 13

How is an AI agent different from a chatbot?

Accepted Answer

A chatbot responds to messages. An AI agent pursues goals. A chatbot is reactive: you send a prompt, it returns a reply, and the interaction ends at that exchange. The chatbot has no persistent internal state between turns beyond the conversation window, calls no external tools, and takes no actions in the world beyond generating text. The original ChatGPT (November 2022) is a chatbot; so is most customer-service software that preceded LLMs. An AI agent is proactive within an objective: given a goal, it plans, calls tools, observes results, and iterates — often over many steps, many minutes, or many hours — until the goal is achieved or the agent gives up. Depending on which tools it has access to, it can browse the web, write and execute code, manage files, send emails, book appointments, and interact directly with software interfaces. Devin (Cognition, March 2024) and OpenAI Operator (January 2025) are agents. The technical distinction is the ReAct loop: reason about the current state → act by calling a tool → observe the result → reason again → repeat. Chatbots complete one turn. Agents run loops. The same underlying LLM can power both — what makes something an agent is the surrounding infrastructure: the planning loop, the tool integrations, and the memory system that persists state across steps. A useful frame: a chatbot is like a very knowledgeable person who only answers questions. An AI agent is like a contractor who takes on a project and reports back when it's done — or when they need guidance.

Question 14

What is the ReAct paper and why does it matter?

Accepted Answer

The ReAct paper — ReAct: Synergizing Reasoning and Acting in Language Models — was posted to arXiv on October 6, 2022 (arXiv:2210.03629) and presented at ICLR 2023. Authors: Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. The paper proposes interleaving reasoning traces ("Thought") with actions ("Act") and observations from the environment, in a single LLM prompt-completion loop. The pattern — think about what to do, do it, observe what happened, think again, repeat — became the architectural template for nearly every LLM agent framework that followed. AutoGPT, BabyAGI, LangChain agents, CrewAI, AutoGen, and modern Agents SDKs all derive from or implement a version of the ReAct pattern. Before ReAct, there was no agreed-upon architectural template for how an LLM-based agent should plan and act. After ReAct, the field had a shared paradigm. It is the single most-cited methodological reference in the modern AI-agent literature and is the conceptual bridge between the frontier LLM capabilities that emerged in 2022 and the autonomous-agent products that appeared in 2023.

Question 15

Was AutoGPT the first AI agent?

Accepted Answer

No, but it is the project that made "AI agent" mainstream. AutoGPT was released on March 30, 2023, two days after Yohei Nakajima's BabyAGI (March 28). AutoGPT became the top-trending repository on GitHub by April 3, 2023, and crossed 100,000 stars within weeks — the fastest-growing open-source project of its era at that scale. Both BabyAGI and AutoGPT were preceded by Adept's ACT-1 demo (September 2022) and the ReAct paper (October 2022). The deeper history of AI agents runs to the Contract Net Protocol (1980), BDI architectures (1987), and the multi-agent systems literature of the 1990s. AutoGPT is best understood as the cultural origin of what the public calls "AI agents" — the project that introduced the concept to a mass audience. It is not the technical origin of the concept, which is decades older.

Question 16

What is the Model Context Protocol (MCP)?

Accepted Answer

The Model Context Protocol (MCP) is an open standard released by Anthropic on November 25, 2024, for connecting AI agents to external tools, data sources, and applications. The core idea: rather than every AI agent implementing its own custom integrations to every tool, MCP provides a standard interface so that any MCP-compatible agent can connect to any MCP-compatible tool without custom code on either side. MCP is analogous to what USB is to hardware peripherals — a universal connector that removes the need for device-specific drivers. Before MCP, connecting an AI agent to a calendar, a database, a web browser, or a code interpreter required custom integration work for each pairing. After MCP, a tool built once as an MCP server is available to any MCP-compatible agent. MCP was rapidly adopted in 2025. OpenAI, Google, Microsoft, and hundreds of third-party tool providers added MCP support, making it the closest thing the agent industry has to a universal connector standard. See the full MCP timeline entry.

Question 17

What is LangChain and why does it matter for AI agents?

Accepted Answer

LangChain is a Python and JavaScript framework for building applications with large language models, created by Harrison Chase and first released in October 2022. It rose to dominance in 2023 as the de facto framework for building LLM agents, packaging the ReAct pattern, tool use, and memory into reusable, composable modules. LangChain's importance to AI agent history is that it productized the ReAct pattern. Before LangChain, building an LLM agent required implementing the planning loop, tool-calling protocol, and memory management from scratch. LangChain provided these as pre-built components, lowering the barrier to agent development dramatically. The resulting explosion of LLM agent projects in 2023 is partly attributable to LangChain making the scaffolding available. In 2024, the LangChain team released LangGraph, a graph-based orchestration framework for multi-agent systems. LangChain and LangGraph together are among the most widely deployed agent frameworks as of 2026.

Question 18

What is Devin and why was it significant?

Accepted Answer

Devin is an autonomous AI software engineer created by Cognition AI, launched on March 12, 2024. It was founded by Scott Wu, Steven Hao, and Walden Yan. Devin operates in a sandboxed environment with a shell, code editor, browser, and persistent workspace — the same tools a human software engineer would use — and is given coding tasks to complete autonomously. Devin's significance is that it was the first widely recognized commercial AI agent product positioned as a worker rather than an assistant. Earlier agent products helped users with tasks; Devin was marketed as taking over the task entirely. This framing — agent as autonomous employee rather than helpful tool — set the template for the wave of "AI employee" products that followed in 2024–2025. Devin's commercial trajectory also validated the agent-product category: ARR grew from approximately $1M (September 2024) to $73M (June 2025), and Cognition raised $400M at a $10.2B valuation in September 2025, with reports of discussions at roughly $25B by April 2026. See the full Devin timeline entry.

Question 19

What is computer use in AI?

Accepted Answer

"Computer use" refers to AI agents that can operate a computer's graphical interface directly — looking at a screen, moving a cursor, clicking, typing, and scrolling — the same way a human would, rather than calling structured APIs. The distinction matters: most AI agents before 2024 interacted with the world through APIs and structured tool calls. Computer use allows an agent to interact with any software a human can use, including software that has no API — legacy enterprise applications, websites that require login, graphical tools, and so on. It effectively removes the need for custom integrations. Anthropic released computer use in public beta on October 22, 2024, as part of an upgraded Claude 3.5 Sonnet. This was the first widely available computer-use capability from a frontier lab. The earlier conceptual precedent is Adept's ACT-1 demo (September 2022). OpenAI launched Operator — its consumer-facing computer-use agent — on January 23, 2025. See Anthropic computer use and OpenAI Operator in the timeline.

Question 20

Who are the leading AI agent companies today?

Accepted Answer

As of May 2026, leading AI-agent companies include: Cognition — Devin, the AI software engineer. Reported to be in discussions at roughly a $25B valuation in April 2026, after raising $400M at $10.2B in September 2025. ARR grew from $1M (Sep 2024) to $73M (Jun 2025). Sierra — Enterprise customer-experience AI agents. $350M at $10B valuation in September 2025, with over $100M enterprise ARR run rate. Founded by Bret Taylor and Clay Bavor. Anthropic — Claude, the leading frontier model for agents, computer use, and the Model Context Protocol. Claude Opus 4.5 (November 2025) is benchmarked specifically on long-horizon agentic tasks. OpenAI — Operator (consumer browser agent), the Agents SDK, and integrated agent capabilities across the ChatGPT product line. Adept — ACT-1 and successor action transformer models; the earliest commercial lab focused on action-taking AI. Manus / Monica — General-purpose autonomous agent (March 2025 launch); reported strong early benchmark performance. Microsoft — AutoGen (multi-agent framework), Microsoft 365 Copilot agents, and deep investment in the agent ecosystem through OpenAI partnership. LangChain — LangChain and LangGraph, among the most widely deployed agent frameworks. See the funding table on the History page for representative investment rounds.

Question 21

What is the BDI architecture?

Accepted Answer

BDI stands for Belief–Desire–Intention, and it is the dominant architectural model for agent software from the late 1980s through the 2000s — and still influential in academic agent programming today. The foundation is philosophical: Michael Bratman's Intention, Plans, and Practical Reason (Harvard University Press, 1987) argued that rational agents maintain three distinct mental states: beliefs about the current state of the world, desires (goals they want to achieve), and intentions (plans they have committed to pursuing). These three structures together govern how a rational agent decides what to do next. AI researchers operationalized BDI into agent programming systems in the late 1980s and 1990s. The most notable formalization is Anand Rao and Michael Georgeff's work, with Rao introducing AgentSpeak(L) in 1996 — a logic-based programming language in which BDI agents can be formally specified and executed. The Jason interpreter (Hübner and Bordini) later provided an open-source platform for AgentSpeak(L)-based multi-agent systems. BDI's influence on modern LLM agents is indirect but real: the conceptual structure of planning (forming intentions), maintaining a world model (beliefs), and pursuing objectives (desires) is structurally similar to how LLM-based agents with planning loops, memory, and goal specifications are designed — though most modern agent builders arrived at this structure independently, via the ReAct paper rather than the BDI literature. See the BDI timeline entry.

Question 22

What is the Contract Net Protocol?

Accepted Answer

The Contract Net Protocol (CNP) is a framework for multi-agent coordination published by Reid G. Smith in IEEE Transactions on Computers (Vol. C-29, No. 12, December 1980). It describes how autonomous computing nodes can negotiate task allocation through a bidding process: a manager node broadcasts a task announcement; contractor nodes submit bids indicating their capability and availability; the manager awards the task to the best bidder; the contractor executes and reports results. The CNP is the founding document of formal multi-agent coordination. It solves a real distributed-systems problem — how do autonomous nodes decide who should do what, without a central controller — using a protocol that is still recognizable in how modern multi-agent LLM systems assign tasks across specialized agents. Smith's 1980 paper introduced the term "contract net" and gave the field of multi-agent systems its first formal, citable coordination algorithm. It is the canonical starting point of the multi-agent systems literature and is still cited in contemporary work. See the Contract Net Protocol timeline entry.

Question 23

Why call this site a museum?

Accepted Answer

Because that is what it is: a curated, dated, sourced archive of how a field came to exist. Museums preserve objects and the context around them; we preserve product launches, papers, demos, and funding rounds, and the context that links them into a lineage. The choice is also a stance: we are not selling an agent. We are not a lab, a startup, or a media company with a financial stake in which version of history is told. We are documenting the record — with sources and dates and acknowledged uncertainties about "firsts" — so that researchers, journalists, builders, and historians have a neutral place to start. See About the Museum for our methodology and source standards.

AI Agent History FAQ — Agentic AI & AI Agent History Frequently Asked Questions

Evidence Standard for FAQ Answers

How Answers Are Sourced

Primary Sources First

How "First" Claims Are Handled

Multiple Definitions, Multiple Valid Dates

How Updates Are Made

Material Corrections and Date Changes

Origins & pioneers explained

Foundational Questions About Origins