Research Timeline

Agentic AI & AI Agents: A Chronological Research Timeline

Last updated: May 9, 2026

This monograph is the timeline maintained by Agentic History — the museum of agentic AI and AI agents. It traces the long arc of the field, from the 1980 Contract Net Protocol and the founding multi-agent-systems literature, through the belief–desire–intention architectures of the late 1980s and the Russell & Norvig synthesis of 1995, into the LLM-tool-use breakthrough of 2022, the AutoGPT and BabyAGI moment of 2023, the rise of computer-use and browser agents in 2024–2025, and the resulting funding wave behind autonomous agent platforms.

Use this page as a citable timeline of AI agents and as a reference for the pioneers of AI agents, the pioneers of multi-agent systems, and the pioneers of modern multi-agent systems. For shorter direct answers — including the full agentic AI history in question-and-answer form — see the FAQ; for the project's research methodology, see About.

Contents
  1. Timeline (Primary)
  2. Funding & Ecosystem (Selected 2024–2025)
  3. Curatorial Notes on Claims
  4. Bibliography / Primary Sources

Timeline

  1. 1956 (Antecedents) The Dartmouth Workshop and the Founding of AI

    The 1956 Dartmouth Summer Research Project on Artificial Intelligence, organized by John McCarthy with Marvin Minsky, Claude Shannon, and Nathaniel Rochester, names the field. Early programs such as the Logic Theorist (Newell, Simon, and Shaw, 1956) demonstrate that machines can carry out goal-directed symbolic reasoning — the conceptual seed of all later "agents."

    Refs: Dartmouth (1956), Logic Theorist, Russell & Norvig (intro).

  2. December 1980 Contract Net Protocol — The First Multi-Agent Coordination Framework

    Reid G. Smith publishes The Contract Net Protocol: High-Level Communication and Control in a Distributed Problem Solver in IEEE Transactions on Computers (Vol. C-29, No. 12, December 1980). The paper formalizes how autonomous nodes — "agents" in everything but name — can negotiate task allocation through a manager–contractor bidding loop. CNP becomes the canonical reference for distributed problem solving and is still cited in modern multi-agent systems literature.

    Refs: Smith (1980), Contract Net Protocol.

  3. 1986 Minsky's Society of Mind — Intelligence as a Multi-Agent System

    Marvin Minsky publishes The Society of Mind (Simon & Schuster, 1986), arguing that intelligence is not produced by a single mechanism but emerges from the interaction of many small, mindless "agents." The book is the most widely cited theoretical foundation for the multi-agent view of cognition and, four decades later, is repeatedly invoked as the conceptual blueprint for AutoGen, CrewAI, and other modern multi-agent frameworks.

    The same year, Rodney Brooks introduces the subsumption architecture, distributing intelligent behavior across layers of simple finite-state machines — a parallel argument that complex behavior can arise from interacting simple agents.

    Refs: Society of Mind, Brooks subsumption.

  4. 1987 BDI — The Belief–Desire–Intention Architecture

    Philosopher Michael Bratman publishes Intention, Plans, and Practical Reason (Harvard University Press, 1987), supplying the philosophical foundation that AI researchers will operationalize as the BDI agent architecture. BDI gives autonomous systems a structured way to maintain beliefs about the world, generate desires (goals), and commit to intentions (plans) — and remains the dominant architecture in academic agent programming languages decades later.

    Refs: Bratman (1987), BDI survey.

  5. 1990s The Multi-Agent Systems Era

    The 1990s are the foundational decade for academic multi-agent systems. Researchers formalize agent communication protocols (KQML, then FIPA-ACL), coordination strategies, coalition formation, and negotiation. Agent-Oriented Programming (Yoav Shoham, 1993) proposes treating agents as a primary unit of software design. The era produces an extensive literature that today's LLM-agent frameworks largely re-derive.

    A significant contribution from this period is AgentSpeak(L), introduced by Dr. Anand Rao in his 1996 paper AgentSpeak(L): BDI Agents Speak Out in a Logical Computable Language. AgentSpeak(L) operationalizes the BDI model in a logic-based programming language, giving developers a formal, executable notation for writing agents with beliefs, goals, and plans. It bridges the gap between the theoretical BDI architecture (Bratman, 1987) and practical agent implementation, and becomes the basis for a family of agent programming languages still in academic and applied use today.

    Refs: Shoham (1993), KQML, FIPA-ACL, Rao — AgentSpeak(L) (1996).

  6. 2005 — ongoing Jason — Open-Source Interpreter for AgentSpeak, Extending Rao's Work

    Jason is an open-source interpreter for an extended version of AgentSpeak(L), implementing its operational semantics and providing a full platform for the development of multi-agent systems. It was brought to our attention by a reader email to the curator (May 2026) as a significant and underrepresented entry in the lineage from BDI theory to deployable agent infrastructure.

    Jason extends Rao's AgentSpeak with practical additions including: communication between agents, annotation-based belief management, custom internal actions, and a rich set of user-customisable features that make it suitable for real multi-agent deployments rather than purely academic demonstrations. The platform is available as open-source software distributed under the GNU LGPL licence.

    Jason is developed by Jomi F. Hübner (University of Santa Catarina) and Rafael H. Bordini (PUCRS, Brazil), building on prior work with Michael Fisher, Joyce Martins, Álvaro Moreira, Renata Vieira, Willem Visser, Mike Wooldridge, and others. It remains one of the most complete and actively maintained BDI agent programming environments, and its lineage — from Bratman's philosophical BDI model (1987), through Rao's AgentSpeak formalization (1996), to Jason's executable open-source platform — represents one of the clearest continuous threads from foundational theory to working multi-agent software.

    Refs: Rao — AgentSpeak(L) (1996), Hübner & Bordini — Jason (jason-lang.github.io), Bratman (1987).

  7. 1995 Russell & Norvig — AI Defined as the Study of Intelligent Agents

    Stuart Russell and Peter Norvig publish the first edition of Artificial Intelligence: A Modern Approach (Prentice Hall, 1995), which reorganizes the entire field around the concept of the rational agent — a system that perceives an environment and acts to maximize expected utility. The textbook becomes the most widely used AI text in the world and cements "agent" as the standard unit of analysis in academic AI for the next three decades.

    Refs: Russell & Norvig (1995).

  8. 2000s Reinforcement Learning, Recommender Agents, and the Practical Turn

    Through the 2000s, machine learning and big data shift agent research from symbolic reasoning to learned behavior. Reinforcement-learning agents, recommender systems, web crawlers, and trading agents prove out the practical model of perceive → decide → act at scale. Sutton and Barto's Reinforcement Learning: An Introduction (1998; 2nd ed. 2018) defines the technical vocabulary that later powers DeepMind's game-playing agents.

    Refs: Sutton & Barto.

  9. December 19, 2013 Deep Reinforcement Learning Agents Reach the Mainstream

    DeepMind publishes Playing Atari with Deep Reinforcement Learning (Mnih et al., arXiv:1312.5602), demonstrating that a single neural-network agent can learn to play seven Atari games at human-comparable levels directly from pixels. The paper kicks off a decade of high-profile agent demonstrations — AlphaGo (2016), AlphaZero (2017), AlphaStar (2019) — that establish that learned agents can master complex environments without hand-crafted rules.

    Refs: Mnih et al. (2013), AlphaGo.

  10. June 11, 2020 The OpenAI API and the Beginning of LLM-as-Agent

    OpenAI releases its API in private beta, making GPT-3 available over HTTP. For the first time, developers can call a general-purpose language model from any program. This is the substrate on which all modern LLM-based AI agents are eventually built: a model that can be invoked, given context, and asked to produce structured output that another program can act on.

    Refs: OpenAI API (2020).

  11. September 14, 2022 Adept's ACT-1 — A Transformer Trained to Take Actions

    Adept AI, founded earlier in 2022 by David Luan and Transformer co-authors Ashish Vaswani and Niki Parmar, demos ACT-1 (Action Transformer) — a model trained to operate web browsers and software interfaces in response to natural-language commands. ACT-1 is the first widely publicized demonstration that a transformer can act as a general digital agent, clicking, typing, and navigating UIs the way a human would. It anticipates by two years the "computer use" capabilities later released by Anthropic and OpenAI.

    Refs: Adept ACT-1 (Sep 2022), Adept Series B (Mar 2023).

  12. October 6, 2022 ReAct — The Paper That Defines the Modern LLM Agent

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao post ReAct: Synergizing Reasoning and Acting in Language Models to arXiv (2210.03629). The paper proposes interleaving reasoning traces ("Thought") with actions ("Act") and observations from the environment — a pattern that becomes the architectural template for nearly every LLM agent framework that follows. ReAct is presented at ICLR 2023 and is the single most-cited methodological reference in the modern AI-agent literature.

    Refs: Yao et al. (2022).

  13. March 1, 2023 ChatGPT and Whisper APIs — The "API Moment" for Agents

    OpenAI releases the ChatGPT API (gpt-3.5-turbo) and the Whisper API at a price point cheap enough to power agent loops. Together with the GPT-4 release on March 14, 2023, this is the inflection point at which building autonomous LLM agents becomes economically and technically possible for individual developers. Within weeks, the open-source agent ecosystem explodes.

    Refs: ChatGPT/Whisper APIs (Mar 2023), GPT-4 release.

  14. March 28, 2023 BabyAGI — The First Widely Shared Autonomous LLM Agent

    Yohei Nakajima publishes Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications and posts the BabyAGI script on GitHub. In roughly 140 lines of Python, BabyAGI demonstrates the canonical autonomous-agent loop: a user-provided objective → task creation → task execution against an LLM → reprioritization based on results → repeat. BabyAGI is the first viral, publicly available implementation of an autonomous LLM agent and is cited in dozens of subsequent academic papers.

    Refs: Nakajima (Mar 28, 2023), BabyAGI (IBM).

  15. March 30, 2023 AutoGPT — The Project That Defines "Autonomous Agent" in the Public Mind

    Toran Bruce Richards, founder of Significant Gravitas Ltd., releases AutoGPT on GitHub. AutoGPT pairs GPT-4 with a self-prompting loop, web browsing, file operations, and code execution to pursue a user-defined goal autonomously. By April 3, 2023, AutoGPT is the top-trending repository on GitHub; within weeks it surpasses 100,000 stars, the fastest-growing open-source project of its era at that scale.

    Why this entry is treated as the popular-discourse "first."

    Earlier projects (BabyAGI, predecessor research notebooks, Adept's ACT-1 demo) shipped first or at the same time, but AutoGPT is the project that introduced the term "autonomous AI agent" to the mainstream and triggered the agent gold rush of April–June 2023. We record both BabyAGI (Mar 28) and AutoGPT (Mar 30) as effectively simultaneous origins of the modern open-source autonomous-agent movement.

    Refs: AutoGPT (Wikipedia), AutoGPT (AI Wiki), Fortune coverage.

  16. 2023 LangChain, AgentGPT, and the Framework Explosion

    Through 2023, LangChain (Harrison Chase, originally released October 2022) becomes the de facto framework for building LLM agents, packaging the ReAct pattern, tool use, and memory into a Python and JavaScript library. AgentGPT, HuggingGPT, CAMEL, and dozens of other frameworks ship in rapid succession. By the end of 2023, "agent" is the default GitHub label for LLM-powered applications.

    Refs: LangChain, AgentGPT.

  17. March 12, 2024 Cognition's Devin — The First "AI Software Engineer" Agent

    Cognition (founded late 2023 by Scott Wu, Steven Hao, and Walden Yan) launches Devin, marketed as the first end-to-end autonomous AI software engineer. Devin operates in a sandboxed environment with a shell, editor, browser, and persistent workspace, and is benchmarked against SWE-bench. The launch sets the template for agent products as long-running, autonomous workers rather than chatbots. Devin's annualized revenue grows from roughly $1M (Sep 2024) to $73M (Jun 2025) and Cognition reaches a $10.2B valuation in September 2025.

    Refs: Devin launch, Cognition $400M / $10.2B, Cognition $25B talks.

  18. Mar–May 2024 Multimodal & Long-Context Models Enable Capable Agents

    Mar 4, 2024. Anthropic announces the Claude 3 family (Haiku, Sonnet, Opus), with strong reasoning and tool-use scores.

    Feb 2024. Google introduces Gemini 1.5 Pro with breakthrough long-context (up to 1M tokens), letting agents ingest entire codebases.

    May 13, 2024. OpenAI unveils GPT-4o, lowering cost and latency for the multimodal voice-and-vision interactions agents increasingly require.

    These models are what make 2024–2025 agent products commercially viable.

    Refs: Claude 3, Gemini 1.5, GPT-4o.

  19. October 22, 2024 Anthropic's Computer Use — Agents That Operate Real Computers

    Anthropic releases computer use in public beta as part of an upgraded Claude 3.5 Sonnet. Claude becomes the first frontier AI model to officially expose a capability for looking at a screen, moving a cursor, clicking, and typing — operating standard human software through vision and input rather than custom APIs. This marks the transition of agents from "tools that call APIs" to "agents that operate the same software humans do."

    Refs: Anthropic computer use.

  20. November 25, 2024 Model Context Protocol — A Standard for Agent–Tool Connections

    Anthropic open-sources the Model Context Protocol (MCP), an open standard for connecting AI agents to external tools, data sources, and applications. MCP is rapidly adopted across the ecosystem in 2025, including by OpenAI, Google, and Microsoft, becoming the closest thing the agent industry has to a universal connector standard.

    Refs: MCP announcement.

  21. January 23, 2025 OpenAI Operator — Browser-Native Agentic AI for Consumers

    OpenAI launches Operator (research preview) for ChatGPT Pro subscribers, built on its Computer-Using Agent (CUA) model. Operator can autonomously navigate the web, fill forms, book travel, and complete shopping tasks. It is the first agent product from a frontier lab targeted at end consumers rather than developers.

    Refs: OpenAI Operator.

  22. March 2025 Manus — General-Purpose Autonomous Agent Goes Viral

    Chinese company Monica releases Manus, an autonomous agent that plans and executes complex multi-step tasks (trip planning, website building, comparative research) end-to-end. Early users report >86% success rates on the GAIA benchmark. Manus drives a renewed wave of public interest in fully autonomous, generalist AI agents.

    Refs: Manus launch coverage.

  23. March 2025 Agent SDKs from the Frontier Labs

    OpenAI releases the Agents SDK and Responses API, productizing the agent loop. Anthropic, Google, and others ship comparable frameworks. By Q2 2025, every major frontier lab offers a first-party SDK explicitly named "agents," and the term replaces "assistants" in most commercial AI marketing.

    Refs: OpenAI Agents SDK.

  24. Q1–Q3 2025 The AI Agent Funding Wave

    According to PitchBook, investment in AI-agent startups reached approximately $8.7 billion in Q1 2025 alone — a roughly 143% year-over-year increase. Cognition (Devin), Sierra (enterprise customer-experience agents), Adept, Manus, and dozens of other agent-first companies raise at multi-billion-dollar valuations through the year. See the funding table below for representative rounds.

    Refs: PitchBook (Q1 2025); see Funding table.

  25. November 24, 2025 Claude Opus 4.5 — Frontier Performance for Long-Horizon Agents

    Anthropic releases Claude Opus 4.5, marketed as the best model for "coding, agents, and computer use," with state-of-the-art performance on real-world software-engineering tasks and sustained 30-minute autonomous coding sessions. The release crystallizes a year-long industry shift: model capability is now measured primarily by how well a model performs as the brain of an agent, not as a single-turn chatbot.

    Refs: Claude Opus 4.5.

Funding & Ecosystem (Selected 2024–2025)

Representative financings of AI-agent companies. Numbers reflect publicly reported rounds and primary sourcing where available.

Date Company / Tool Amount Valuation Notes Source
Mar 2023 Adept AI $350M Series B ~$1B ACT-1 maker; founded by Transformer co-authors and ex-OpenAI/Google leads. Wikitia / coverage
Oct 2023 AutoGPT (Significant Gravitas) $12M — Led by Redpoint and GitHub Ventures. AutoGPT funding
Mar 2024 Cognition (Devin) $21M Series A ~$350M Founders Fund; Devin launched the same month. Cognition (Sacra)
Oct 2024 Sierra $175M $4.5B Greenoaks; enterprise customer-experience AI agents (Bret Taylor, Clay Bavor). Sierra (Sacra / Axios)
Mar 2025 Cognition (Devin) (round) ~$4B Reported valuation jump on Devin's accelerating ARR. TechCrunch / Bloomberg
Sep 2025 Sierra $350M $10B Greenoaks; >$100M enterprise ARR run rate. SiliconANGLE / Axios
Sep 8, 2025 Cognition (Devin) $400M $10.2B Founders Fund; Devin ARR $1M (Sep 2024) → $73M (Jun 2025); follows Windsurf acquisition. Cognition blog / TechCrunch
Apr 2026 Cognition (Devin) (in talks) ~$25B (reported) Bloomberg / SiliconANGLE: discussions for round at ~$25B post-Windsurf. SiliconANGLE / TFN

Curatorial Notes on Claims

Bibliography / Primary Sources

  1. McCarthy, J., Minsky, M., Rochester, N., Shannon, C. (1955/1956). "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence."
  2. Newell, A., Simon, H., Shaw, J. C. (1956). The Logic Theorist — primary discussion in subsequent Newell & Simon work.
  3. Smith, R. G. (1980). "The Contract Net Protocol: High-Level Communication and Control in a Distributed Problem Solver." IEEE Transactions on Computers, Vol. C-29, No. 12, December 1980.
  4. Wikipedia: Contract Net Protocol.
  5. Minsky, M. (1986). The Society of Mind. Simon & Schuster.
  6. Brooks, R. (1986). "A Robust Layered Control System for a Mobile Robot." IEEE Journal of Robotics and Automation.
  7. Bratman, M. (1987). Intention, Plans, and Practical Reason. Harvard University Press.
  8. Rao, A. S., & Georgeff, M. P. (1995). "BDI Agents: From Theory to Practice."
  9. Shoham, Y. (1993). "Agent-Oriented Programming." Artificial Intelligence 60(1).
  10. Finin, T., et al. (1994). "KQML as an Agent Communication Language."
  11. FIPA (1997+). FIPA Agent Communication Language specifications.
  12. Rao, A. S. (1996). "AgentSpeak(L): BDI Agents Speak Out in a Logical Computable Language." Proceedings of the 7th European Workshop on Modelling Autonomous Agents in a Multi-Agent World (MAAMAW-96).
  13. Hübner, J. F. & Bordini, R. H. Jason: An open-source interpreter for an extended version of AgentSpeak. Developed with contributions from Michael Fisher, Joyce Martins, Álvaro Moreira, Renata Vieira, Willem Visser, Mike Wooldridge, and others. Available under GNU LGPL. jason-lang.github.io.
  14. Russell, S. & Norvig, P. (1995, 1st ed.). Artificial Intelligence: A Modern Approach. Prentice Hall.
  15. Sutton, R. S. & Barto, A. G. (1998; 2nd ed. 2018). Reinforcement Learning: An Introduction. MIT Press.
  16. Mnih, V., et al. (2013). "Playing Atari with Deep Reinforcement Learning." arXiv:1312.5602.
  17. Silver, D., et al. (2016). "Mastering the game of Go with deep neural networks and tree search." Nature.
  18. OpenAI (June 11, 2020). "OpenAI API."
  19. Adept AI (Sep 14, 2022). "ACT-1: Transformer for Actions" (adept.ai/act).
  20. Adept AI Series B coverage (March 2023): $350M at ~$1B valuation.
  21. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2022/2023). "ReAct: Synergizing Reasoning and Acting in Language Models." arXiv:2210.03629; ICLR 2023.
  22. OpenAI (March 1, 2023). "Introducing ChatGPT and Whisper APIs."
  23. OpenAI (March 14, 2023). "GPT-4."
  24. Nakajima, Y. (March 28, 2023). "Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications."
  25. IBM Think: "What is BabyAGI?" — secondary reference.
  26. Wikipedia: AutoGPT (release: March 30, 2023, by Toran Bruce Richards / Significant Gravitas Ltd.).
  27. AI Wiki: AutoGPT.
  28. AutoGPT / Significant Gravitas $12M round (October 2023; Redpoint, GitHub Ventures).
  29. Roose / Fortune (April 2023). Coverage of AutoGPT, BabyAGI, AgentGPT.
  30. LangChain (Harrison Chase). Initial release October 2022; rapid 2023 adoption as the dominant LLM-agent framework.
  31. AgentGPT (2023). Browser-based autonomous agent built on the AutoGPT pattern.
  32. Cognition (March 12, 2024). Devin — the first AI software engineer.
  33. Cognition blog (Sep 8, 2025), TechCrunch (Sep 8, 2025), Sacra: Cognition $400M at $10.2B valuation; Devin ARR growth $1M → $73M.
  34. SiliconANGLE / Tech Funding News (April 2026). Cognition in talks to raise at ~$25B valuation.
  35. Sierra (Sep 2025). $350M at $10B; founders Bret Taylor, Clay Bavor (Axios, SiliconANGLE, Sacra).
  36. Anthropic (March 4, 2024). "Introducing the next generation of Claude" (Claude 3 family).
  37. Google (February 2024). "Gemini 1.5: Our next-generation model."
  38. OpenAI (May 13, 2024). "Hello GPT-4o."
  39. Anthropic (October 22, 2024). "Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku."
  40. Anthropic (November 25, 2024). "Introducing the Model Context Protocol."
  41. OpenAI (January 23, 2025). Operator research preview launch.
  42. Manus / Monica (March 2025). Public launch coverage; GAIA benchmark reporting.
  43. OpenAI (March 2025). Agents SDK and Responses API.
  44. PitchBook (Q1 2025). AI-agent startup funding totals: ~$8.7B in Q1 2025.
  45. Anthropic (November 24, 2025). "Introducing Claude Opus 4.5."

Continue exploring: What is an AI agent? · About the Museum · FAQ