History of Agents

Overview

The concept of agents has evolved from thought experiments in the 1950s through multiple paradigm shifts spanning symbolic AI, behaviorism, hybrid architectures, the internet era, and the large language model era. This article traces the complete timeline of agent development and analyzes the technical contributions and historical significance of each milestone.

Development Timeline

timeline
    title History of Agents (1950-2025)
    section Early Exploration (1950-1970)
        1950 : Turing Test - Defining "Machine Intelligence"
        1956 : Dartmouth Conference - Birth of AI as a Discipline
        1966 : ELIZA - First Conversational Agent
        1969 : Shakey - First General-Purpose Mobile Robot
        1971 : SHRDLU - Natural Language Understanding System
    section Knowledge Era (1970-1990)
        1972 : MYCIN - Expert System Pioneer
        1980s : Golden Age of Expert Systems
        1987 : BDI Model - Theoretical Foundation for Rational Agents
        1986 : Brooks Subsumption Architecture - Behavior-Based Agents
    section Multi-Agent Era (1990-2010)
        1994 : Softbots - Software Agent Concept
        1995 : KQML/ACL - Agent Communication Languages
        2002 : AAMAS - Premier Multi-Agent Systems Conference Founded
        2003 : SOAR 9 - Cognitive Architecture Matures
    section Deep Learning Era (2010-2020)
        2011 : Siri - First Mainstream Voice Assistant
        2014 : Alexa/Cortana - Smart Assistants Go Mainstream
        2016 : AlphaGo - Deep RL Milestone
        2017 : Transformer - Attention Mechanism Revolution
    section LLM Agent Era (2020-2025)
        2022 : ChatGPT - LLM Conversational Agent
        2023 : AutoGPT/BabyAGI - Autonomous Agents
        2023 : GPT-4 Function Calling
        2024 : Claude Code - Programming Agent
        2025 : OpenAI Operator - Browser Agent

1. Theoretical Foundations (1950-1966)

1.1 The Turing Test (1950)

Alan Turing proposed the famous "Imitation Game" in his paper Computing Machinery and Intelligence, laying the philosophical foundations for agent research.

Key Contributions:

Proposed an operational criterion for determining machine intelligence
Transformed "Can machines think?" into a testable behavioral question
Foresaw the core challenges of conversational agents

1.2 The Dartmouth Conference (1956)

Organized by McCarthy, Minsky, Shannon, Rochester, and others, this conference formally established "Artificial Intelligence" as an academic discipline.

Key Contributions:

Defined the fundamental goals and methodology of AI research
Gave rise to the Symbolic AI research paradigm
Provided the disciplinary framework for subsequent agent research

1.3 ELIZA (1966)

ELIZA, developed by Joseph Weizenbaum at MIT, was the first program capable of engaging in natural language conversation with humans.

Key Contributions:

Demonstrated that simple pattern matching could produce the illusion of "intelligence"
First revealed the human tendency to anthropomorphize machines (the ELIZA effect)
Although lacking genuine "understanding," it pioneered conversational agents

2. The Peak of Symbolic AI (1969-1990)

2.1 Shakey (1969)

Shakey, developed by SRI International, was the first general-purpose mobile robot capable of autonomous reasoning and action planning.

Key Contributions:

Integrated the complete agent loop of perception, planning, and execution
Used STRIPS (Stanford Research Institute Problem Solver) as its planner
Demonstrated that symbolic reasoning could drive actions in the physical world

2.2 SHRDLU (1971)

Terry Winograd's SHRDLU could understand natural language instructions and manipulate objects in a virtual blocks world.

Key Contributions:

Demonstrated the integration of language understanding and action execution
Proved that "deep understanding" was possible within restricted domains
Revealed the fundamental difficulty of open-world understanding

2.3 The Expert System Era (1970s-1980s)

Expert systems such as MYCIN (medical diagnosis), DENDRAL (chemical analysis), and R1/XCON (computer configuration) demonstrated the practical value of knowledge-driven agents in specific domains.

Key Contributions:

Introduced the architectural pattern of Knowledge Base + Inference Engine
Uncertainty reasoning: MYCIN introduced Certainty Factors
Achieved the first commercial deployment of AI systems

2.4 The BDI Model (1987)

The Belief-Desire-Intention model proposed by Bratman provided a rigorous philosophical and computational foundation for rational agents.

Key Contributions:

Formally defined the mental states of agents
Provided a rational decision-making framework from beliefs to actions
Gave rise to implementations such as PRS and AgentSpeak(L)

Cross-Reference

For a detailed formal description of the BDI model, see BDI Model.

2.5 Brooks' Subsumption Architecture (1986)

The Subsumption Architecture proposed by Rodney Brooks challenged the dominance of symbolic AI.

Key Contributions:

Proposed that "the world is its own best model" -- no internal representation needed
Behavior layering: lower-level behaviors take priority; higher-level behaviors can suppress lower ones
Founded behavior-based agent research

3. The Multi-Agent and Internet Era (1994-2010)

3.1 Softbots (1994)

Etzioni and Weld proposed the concept of "software agents" (Softbots) -- programs that autonomously execute tasks in digital environments.

Key Contributions:

Extended the agent concept from the physical world to the digital world
Foresaw later web crawlers, automation scripts, and RPA
Raised safety and trust issues for software agents

3.2 Agent Communication Languages (1995-2002)

KQML (Knowledge Query and Manipulation Language) and FIPA ACL provided standardized protocols for inter-agent communication.

Key Contributions:

Defined the application of speech acts in agent communication
Established standards for message formats, interaction protocols, and ontologies
Laid the foundation for multi-agent system interoperability

3.3 The AAMAS Conference (2002)

The founding of the International Conference on Autonomous Agents and Multiagent Systems marked multi-agent research as an independent field.

4. The Deep Learning and Voice Assistant Era (2011-2020)

4.1 Siri (2011)

Apple's Siri brought intelligent assistants into the mainstream consumer market.

Key Contributions:

Integrated the complete pipeline of speech recognition, natural language understanding, and task execution
Proved that agents could serve as mass consumer products
Launched the competition with Alexa (2014), Google Assistant (2016), and Cortana (2014)

4.2 AlphaGo (2016)

DeepMind's AlphaGo defeated world champion Lee Sedol in the game of Go.

Key Contributions:

Demonstrated superhuman capabilities of deep reinforcement learning in complex decision tasks
Combined Monte Carlo Tree Search (MCTS) with deep neural networks
Inspired subsequent research on applying RL to broader agent tasks

4.3 Transformer (2017)

The Transformer architecture proposed by Vaswani et al. fundamentally transformed NLP, laying the technical foundation for later LLM agents.

5. The LLM Agent Era (2022-2025)

5.1 ChatGPT (2022)

OpenAI's ChatGPT showcased the powerful capabilities of LLMs as conversational agents.

Key Contributions:

RLHF-aligned LLM achieved fluent, helpful multi-turn conversation for the first time
Brought AI agents from research labs to hundreds of millions of users
Triggered an explosive growth in LLM agent research

5.2 AutoGPT and BabyAGI (March 2023)

Significant Gravitas' AutoGPT and Yohei Nakajima's BabyAGI were among the first autonomous LLM agents.

Key Contributions:

Demonstrated that LLMs could autonomously decompose tasks, plan steps, and execute them
Introduced mechanisms such as task queues, memory storage, and self-reflection
Although limited in practical reliability, they sparked the imagination of the entire field

5.3 GPT-4 Function Calling (June 2023)

OpenAI introduced structured function calling capabilities for GPT-4.

Key Contributions:

Elevated tool use from prompt engineering to a native capability
Standardized the interaction interface between LLMs and external tools
Provided infrastructure for subsequent agent frameworks (LangChain, LangGraph)

5.4 Claude Code (2024)

Claude Code, released by Anthropic, was one of the first production-grade programming agents.

Key Contributions:

Demonstrated that LLM agents can work reliably on real software engineering tasks
Integrated a complete workflow of code understanding, file operations, and test execution
Showcased the Human-in-the-Loop interaction pattern

5.5 OpenAI Operator (January 2025)

OpenAI launched Operator, an agent capable of autonomously browsing the web and completing tasks.

Key Contributions:

First commercial autonomous browser agent
Capable of handling complex tasks such as web navigation, form filling, and shopping
Drove the paradigm shift of agents from "conversation" to "action"

Core Threads of Technological Evolution

Era	Core Paradigm	Knowledge Source	Reasoning Method	Action Capability
Symbolic AI (1950-1990)	Rules + Logic	Manual encoding	Deductive reasoning	Restricted environments
Behaviorism (1986-2000)	Reactive	Environmental interaction	No explicit reasoning	Physical world
Statistical Learning (2000-2020)	Data-driven	Labeled data	Pattern matching	Classification/Prediction
LLM Era (2020-)	Language-driven	Pre-training corpora	Analogical reasoning + CoT	Tool calling + Autonomous action

Key Observations

From specialized to general: Agents evolved from specific tasks (MYCIN's medical diagnosis) to general capabilities (GPT-4's multi-domain tasks)
From explicit to implicit knowledge: From manually coded rules to automatic learning from data
From passive to autonomous: From user-driven Q&A to agents that proactively plan and execute
From unimodal to multimodal: From pure text to integrating vision, speech, code, and other modalities
From closed to open environments: From virtual blocks worlds to real web pages, file systems, and the physical world

References

Turing, A.M. (1950). Computing Machinery and Intelligence. Mind, 59(236), 433-460.
Weizenbaum, J. (1966). ELIZA - A Computer Program For the Study of Natural Language Communication Between Man And Machine. CACM, 9(1), 36-45.
Nilsson, N.J. (1984). Shakey the Robot. SRI International Technical Note 323.
Bratman, M.E. (1987). Intention, Plans, and Practical Reason. Harvard University Press.
Brooks, R.A. (1986). A Robust Layered Control System For a Mobile Robot. IEEE Journal on Robotics and Automation, 2(1), 14-23.
Etzioni, O. & Weld, D. (1994). A Softbot-based Interface to the Internet. CACM, 37(7), 72-76.
Vaswani, A. et al. (2017). Attention Is All You Need. NeurIPS 2017.
Significant Gravitas. (2023). AutoGPT. GitHub Repository.
OpenAI. (2023). GPT-4 Technical Report. arXiv:2303.08774.

History of Agents

Overview

Development Timeline

1. Theoretical Foundations (1950-1966)

1.1 The Turing Test (1950)

1.2 The Dartmouth Conference (1956)

1.3 ELIZA (1966)

2. The Peak of Symbolic AI (1969-1990)

2.1 Shakey (1969)

2.2 SHRDLU (1971)

2.3 The Expert System Era (1970s-1980s)

2.4 The BDI Model (1987)

2.5 Brooks' Subsumption Architecture (1986)

3. The Multi-Agent and Internet Era (1994-2010)

3.1 Softbots (1994)

3.2 Agent Communication Languages (1995-2002)

3.3 The AAMAS Conference (2002)

4. The Deep Learning and Voice Assistant Era (2011-2020)

4.1 Siri (2011)

4.2 AlphaGo (2016)

4.3 Transformer (2017)

5. The LLM Agent Era (2022-2025)

5.1 ChatGPT (2022)

5.2 AutoGPT and BabyAGI (March 2023)

5.3 GPT-4 Function Calling (June 2023)

5.4 Claude Code (2024)

5.5 OpenAI Operator (January 2025)

Core Threads of Technological Evolution

Key Observations

References

评论 #