AI Agents History Timeline (1950–2026)

Multi-agent orchestration goes enterprise — and gets regulated

Innovation: 2026 is widely described as the “year of multi-agent systems”, as organisations move from single-agent pilots to fleets of specialised agents coordinated by orchestration layers. Technical breakthrough: two open standards matured — Anthropic’s Model Context Protocol (MCP) for agent-to-tool access and Google’s Agent2Agent (A2A) for agent-to-agent collaboration — adopted across Anthropic, OpenAI, Google, Microsoft and AWS. MCP published its 2026-07-28 specification on 28 July 2026, making the protocol stateless at the transport layer — dropping session IDs and the initialisation handshake so any server instance behind a load balancer can answer a request — while adding multi round-trip requests, header-based routing, cacheable list results and hardened authorisation, and moving long-running Tasks out of the core into a formal extension. A2A, donated by Google to the Linux Foundation in June 2025, marked its first year in April 2026 with a stable v1.0 specification, cryptographically signed Agent Cards, more than 150 supporting organisations and integration into Azure AI Foundry, Copilot Studio, Amazon Bedrock AgentCore and Google Cloud. Product launches: agentic models and platforms reached production — OpenAI’s GPT-5.5 (April 2026) added native computer use and multi-step task execution, Anthropic’s Claude Cowork agent expanded to web and mobile (July 2026), and Microsoft’s Foundry Agent Service became generally available (July 2026) with sandboxed, multi-agent orchestration. OpenAI also agreed to acquire the German cloud-sandbox startup Ona (formerly Gitpod), announced on 11 June 2026, so that Codex agents can keep working on a task unattended inside a customer’s own cloud. In late July and early August 2026 two further shifts stood out: AWS renamed Amazon Bedrock Agents (its November 2023 launch) to Bedrock Agents Classic, closing it to new customers on 30 July 2026 and steering builders to the newer Bedrock AgentCore platform; and Google’s Gemini Spark agent began controlling a user’s actual desktop Chrome browser — using logged-in sessions and saved passwords, with permission prompts and payments handed back to the user — in a US rollout from 3 August 2026. Google’s agentic-calling feature, which has Search dial local businesses on a shopper’s behalf to check price and availability, completed its US-wide rollout by mid-August 2026, and Anthropic’s Claude connectors directory passed 950 MCP servers used by millions of people daily. Investors also backed agent-security infrastructure: data-security firm Cyera agreed on 28 July 2026 to acquire identity-security startup Oasis Security for about $1 billion to govern the access and behaviour of proliferating AI agents. In late August 2026, agent governance kept consolidating: the Linux Foundation’s Agent2Agent (A2A) protocol moved from a standalone project into a hosted project of the newly formed Agentic AI Foundation (AAIF) — the neutral body launched in December 2025 that also anchors Anthropic’s MCP, Block’s Goose and OpenAI’s AGENTS.md — putting the two leading agent standards under one governance umbrella, as reported around 17 August 2026. Days later, on 20 August 2026, Google began bundling its Antigravity AI coding-agent platform into eligible Gemini Enterprise subscriptions, adding centralised license pooling, workspace/browser/MCP-server access policies and audit logging for IT teams, plus new IDE extensions so developers can run Antigravity agents inside VS Code, Visual Studio, JetBrains or Zed rather than only its standalone editor. Framework security also drew scrutiny: Check Point Research disclosed 11 vulnerabilities across widely used agent frameworks — including LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework and Google’s ADK — in research published in early August 2026, including a LangGraph exploit chain (three assigned CVEs) that could grant remote code execution on self-hosted checkpoint servers via SQLite or Redis injection, a reminder that fast-maturing orchestration layers still carry decades-old bug classes like insecure deserialization and injection. On 26 August 2026 Salesforce and Anthropic announced Claudeforce, an expanded partnership that makes Claude a reasoning option inside Agentforce’s Atlas Reasoning Engine and ships “Salesforce in Claude”, a plugin with 37 prebuilt sales skills letting sellers and agents act on live CRM data with governed actions from inside Claude — available to pilot customers immediately, with open beta planned for September 2026, per the companies’ joint press release.

Industry impact (analyst estimates): Gartner projects up to 40% of enterprise applications will embed task-specific AI agents by end-2026, up from under 5% in 2025, with the AI-agent market estimated near $10.8 billion. IDC has projected enterprises will collectively run more than a billion agents by 2029. Gartner also warned of widespread “agentwashing”, estimating only a small fraction — around 130 — of the thousands of vendors claiming agentic AI genuinely deliver it.

Regulation: 2026 produced the first rules aimed at agents as such. On 8 May 2026 China’s Cyberspace Administration, together with the National Development and Reform Commission and the Ministry of Industry and Information Technology, issued Implementation Opinions on the Standardised Application and Innovative Development of Intelligent Agents — the first national framework dedicated to agents — which sorts an agent’s decisions into three authority tiers (human-only, user-approved, agent-autonomous) and imposes filing, testing and review duties in sensitive sectors such as healthcare, transport, media and public safety. In the European Union, the AI Act‘s Article 50 transparency duties — disclosing that a person is interacting with an AI system and marking AI-generated content — became applicable on 2 August 2026, alongside the Commission’s enforcement powers and penalties for general-purpose AI models. A Digital Omnibus amendment adopted shortly before postponed most high-risk obligations to 2 December 2027 (Annex III) and 2 August 2028 (Annex I).

Browser agents, agent SDKs and agent-to-agent protocols

Product launches: OpenAI released Operator (23 January 2025), a browser agent powered by its Computer-Using Agent (CUA) model, and later an Agents SDK and Responses API for building production agents. Technical breakthrough: Google introduced the Agent2Agent (A2A) protocol on 9 April 2025 with 50-plus launch partners, standardising how agents discover and delegate tasks to one another; Google donated the protocol to the Linux Foundation in June 2025, placing it under neutral governance. New general-purpose agents such as Manus also drew attention.

Agentic AI’s breakout year

Product launches: Cognition unveiled Devin (March 2024), billed as the first autonomous “AI software engineer”; OpenAI released the o1 reasoning model (September 2024) that “thinks” step by step; and Anthropic shipped computer use for Claude (October 2024), letting an agent operate a screen. Technical breakthrough: Anthropic also introduced the Model Context Protocol (MCP) in November 2024. Enterprise platforms such as Salesforce Agentforce launched, and frameworks like LangGraph and CrewAI matured.

The autonomous-agent explosion

Product & research: OpenAI’s GPT-4 (March 2023) supercharged a wave of autonomous agents. AutoGPT (30 March 2023) and BabyAGI (April 2023) showed an LLM looping to pursue goals on its own, going viral. Stanford’s Generative Agents paper simulated believable behaviour in a sandbox town, NVIDIA’s Voyager played Minecraft autonomously, OpenAI added function calling (June 2023), and Microsoft released the multi-agent framework AutoGen.

ChatGPT, reasoning and tool use

Research: two 2022 papers laid the intellectual foundation for LLM agents — chain-of-thought prompting (reasoning in steps) and ReAct (interleaving reasoning with actions and tool use). Product launch: OpenAI released ChatGPT on 30 November 2022, built on instruction tuning and reinforcement learning from human feedback (RLHF), reaching an estimated 100 million users within two months and igniting the generative-AI era.

Foundation models and scale

Research: OpenAI’s GPT-3 (2020) showed that scaling language models unlocked few-shot learning — performing new tasks from a prompt alone. In 2021 Stanford researchers coined the term “foundation models” for these general-purpose systems, and GitHub Copilot (built on OpenAI Codex) brought AI code generation to developers. These general models became the substrate on which later agents were built.

The transformer revolution

Research: Google’s 2017 paper “Attention Is All You Need” introduced the transformer architecture, the basis of virtually all modern language models. It was followed by GPT-1 and BERT (2018) and GPT-2 (2019), which demonstrated powerful transfer learning. In parallel, DeepMind’s AlphaZero (2017) mastered chess and Go from self-play, advancing reinforcement-learning agents.

Deep reinforcement learning and virtual assistants

Research: DeepMind’s Deep Q-Network (DQN) (2013, Nature 2015) learned to play Atari games from raw pixels, launching deep reinforcement learning; AlphaGo then defeated Go champion Lee Sedol in March 2016. Products: consumer “agents” reached the mainstream with Apple Siri (2011), IBM Watson winning Jeopardy (2011), and Amazon Alexa (2014) — goal-directed assistants, if far narrower than today’s.

Multi-agent systems and the semantic web

Research: the 2000s matured multi-agent systems (MAS) — multiple autonomous agents coordinating, negotiating or competing — used in logistics, simulation and trading. Tim Berners-Lee’s vision of the Semantic Web (2001) imagined software agents traversing machine-readable data, while recommendation and search “softbots” spread across the early web.

Intelligent software agents and the rational-agent model

Research: the 1990s defined the modern intelligent agent. Pattie Maes pioneered autonomous software agents at MIT; Wooldridge and Jennings published an influential 1995 survey of agent theory; and Stuart Russell and Peter Norvig’s 1995 textbook Artificial Intelligence: A Modern Approach framed AI itself around the rational agent — a system that perceives via sensors and acts via actuators to maximise expected performance. Agent standards (FIPA) and the belief-desire-intention (BDI) architecture also emerged.

Expert systems and the society of mind

Research & industry: the 1980s were the era of expert systems — knowledge-based programs like MYCIN and Digital’s XCON that encoded human expertise as rules to make decisions, powering the first commercial AI boom. Marvin Minsky’s The Society of Mind (1986) argued that intelligence emerges from many simple interacting agents, and Rodney Brooks’ subsumption architecture reimagined robots as reactive, behaviour-based agents.

The birth of artificial intelligence

Foundational research: in 1950 Alan Turing published Computing Machinery and Intelligence, proposing the imitation game now known as the Turing Test. In 1956 the Dartmouth workshop, led by John McCarthy (who coined “artificial intelligence”) with Marvin Minsky, Claude Shannon and others, founded AI as a field. Early milestones included the Logic Theorist (1956), ELIZA (1966), the first chatbot, and SHRDLU (1970), an early language-understanding agent.