# History & Timeline

From Eliza to autonomous agents — the evolution of agentic AI.

---

## Timeline

### 1960s-1990s: Foundations
| Year | Milestone | Significance |
|---|---|---|
| **1966** | **ELIZA** (Weizenbaum) | First chatbot; pattern matching; showed anthropomorphism |
| **1972** | **PARRY** (Colby) | Simulated paranoid schizophrenia; more sophisticated |
| **1980s** | **Expert Systems** (MYCIN, DENDRAL) | Rule-based reasoning; knowledge engineering |
| **1990s** | **BDI Agents** (Rao & Georgeff) | Beliefs-Desires-Intentions formalism |
| **1995** | **FIPA** founded | Agent communication standards (ACL) |
| **1997** | **IBM Deep Blue** | Symbolic search beats Kasparov |

### 2000s-2010s: Pre-LLM Agents
| Year | Milestone | Significance |
|---|---|---|
| **2000s** | **Multi-Agent Systems** (Wooldridge, Jennings) | Coordination, negotiation, organizations |
| **2006** | **SOAR/ ACT-R** cognitive architectures | Unified theories of cognition |
| **2010s** | **Dialogue Systems** (Rasa, Dialogflow) | Intent classification + slot filling |
| **2015** | **Deep Q-Networks** (Mnih et al.) | Deep RL for Atari; neural agents |
| **2016** | **AlphaGo** (Silver et al.) | MCTS + neural nets; superhuman Go |
| **2017** | **Transformers** (Vaswani et al.) | Foundation for all modern LLMs |
| **2018** | **BERT/GPT** | Pretraining + fine-tuning paradigm |

### 2020-2022: LLM Emergence
| Year | Milestone | Significance |
|---|---|---|
| **2020** | **GPT-3** (Brown et al.) | 175B params; few-shot learning |
| **2021** | **Codex** (Chen et al.) | Code generation; GitHub Copilot |
| **2021** | **Chain-of-Thought** (Wei et al.) | "Let's think step by step" |
| **2022** | **InstructGPT** (Ouyang et al.) | RLHF alignment; ChatGPT precursor |
| **2022** | **ReAct** (Yao et al.) | Reasoning + acting loop |
| **2022** | **ChatGPT** (Nov 2022) | Mainstream LLM adoption |

### 2023: Agent Explosion
| Month | Milestone | Significance |
|---|---|---|
| **Mar** | **AutoGPT** (Significant Gravitas) | Viral autonomous agent; recursive self-prompting |
| **Mar** | **BabyAGI** (Yohei Nakajima) | Task management + execution loop |
| **Mar** | **AgentGPT** | Web UI for autonomous agents |
| **Apr** | **LangChain Agents** | Tool use, memory, planning abstractions |
| **May** | **AutoGen** (Microsoft) | Multi-agent conversations, GroupChat |
| **Jun** | **Generative Agents** (Park et al., Stanford) | Sims-like agents with memory/reflection |
| **Jul** | **MetaGPT** (Li et al.) | Software company simulation |
| **Aug** | **CrewAI** | Role-based, process-driven crews |
| **Sep** | **OpenAI Assistants API** | Threads, tools, retrieval, code interpreter |
| **Oct** | **LangGraph** (LangChain) | Stateful graphs, checkpointing, human-in-loop |
| **Nov** | **OpenAI GPT-4 Turbo** | 128k context, function calling improvements |

### 2024: Production & Specialization
| Month | Milestone | Significance |
|---|---|---|
| **Jan** | **SWE-bench** (Yang et al.) | Real GitHub issues benchmark |
| **Feb** | **Devin** (Cognition) | "First AI software engineer" demo |
| **Mar** | **Claude 3** (Anthropic) | 200k context, tool use, vision |
| **Apr** | **Llama 3** (Meta) | Open weights, strong base models |
| **May** | **GPT-4o** (OpenAI) | Native multimodal, real-time audio |
| **Jun** | **OpenAI Agents SDK** | Handoffs, guardrails, tracing |
| **Jul** | **MCP** (Anthropic) | Model Context Protocol standard |
| **Aug** | **Gemini 1.5 Pro** | 1M/2M context, long-context RAG |
| **Sep** | **o1/o1-mini** (OpenAI) | Reasoning models, RL on CoT |
| **Oct** | **Claude 3.5 Sonnet** | Computer use (desktop control) |
| **Nov** | **o1-pro** / **o3** announced | Scaling test-time compute |

### 2025+: Frontiers (Projected)
| Area | Emerging Directions |
|---|---|
| **Test-Time Compute Scaling** | o1-style reasoning, MCTS, self-consistency at inference |
| **Agent-Computer Interface** | Full OS control (Claude Computer Use, OpenInterpreter) |
| **Self-Improving Agents** | RLAIF, constitutional AI, evolutionary agents |
| **Agent Societies** | Civilization sims, economies, governance |
| **Formal Verification** | Proof-carrying agents, model checking |
| **Neurosymbolic Integration** | Logic + neural, program synthesis, SAT/SMT |
| **Lifelong Learning** | Continual adaptation without forgetting |
| **Multimodal Native** | Unified tokenizers for text/image/audio/video |

---

## Key Paradigm Shifts

| Era | Paradigm | Key Insight |
|---|---|---|
| **1960-2000** | Symbolic AI | Hand-coded rules, logic, planning |
| **2000-2017** | Statistical ML | Supervised learning, feature engineering |
| **2017-2022** | Pretraining + Fine-tuning | Transformers, scaling laws, few-shot |
| **2022-2023** | Prompting + CoT | LLMs as reasoners, not just completers |
| **2023-2024** | Tool Use + Memory | LLMs as agents acting in environments |
| **2024-2025** | Reasoning Models | Test-time compute scaling (o1, o3) |
| **2025+** | Autonomous Systems | Self-improving, verified, lifelong agents |

---

## Influential Papers (Chronological)

| Year | Paper | Authors | Impact |
|---|---|---|---|
| 2017 | Attention Is All You Need | Vaswani et al. | Transformer architecture |
| 2020 | Language Models are Few-Shot Learners | Brown et al. | GPT-3, scaling laws |
| 2021 | Evaluating Large Language Models Trained on Code | Chen et al. | Codex, code generation |
| 2022 | Chain-of-Thought Prompting Elicits Reasoning | Wei et al. | CoT paradigm |
| 2022 | ReAct: Synergizing Reasoning and Acting | Yao et al. | Agent pattern |
| 2022 | Training Language Models to Follow Instructions | Ouyang et al. | RLHF/InstructGPT |
| 2023 | Generative Agents: Interactive Simulacra | Park et al. | Memory, reflection, planning |
| 2023 | AutoGen: Enabling Next-Gen LLM Applications | Wu et al. | Multi-agent framework |
| 2023 | MetaGPT: Meta Programming for Software Dev | Li et al. | Multi-agent software teams |
| 2024 | SWE-bench: Can LMs Resolve Real GitHub Issues? | Yang et al. | Coding agent benchmark |
| 2024 | DeepSeek-V2: Strong, Economical MoE | DeepSeek | MLA, efficient MoE |
| 2024 | Scaling Test-Time Compute | OpenAI | o1 paradigm shift |

---

## Framework Genealogy

```
LangChain (2022)
├── LangGraph (2023) ← Stateful graphs
├── LangServe (2023) ← Deployment
└── LangSmith (2023) ← Observability

AutoGen (2023, Microsoft)
├── AutoGen Studio (2024) ← No-code UI
├── Magentic-One (2024) ← Generalist multi-agent
└── AgentEval (2024) ← Evaluation

CrewAI (2023)
├── CrewAI Enterprise (2024)
└── CrewAI Flows (2024) ← Structured processes

OpenAI
├── Assistants API (2023)
├── Function Calling (2023)
├── Agents SDK (2024) ← Handoffs, guardrails
└── Computer Use (2024) ← Desktop control

Anthropic
├── MCP (2024) ← Model Context Protocol
├── Computer Use (2024)
└── Constitutional AI (2022)
```

---

## Related Notes

- [[Foundations/LLM-Fundamentals.md]] — Technical foundations
- [[Foundations/Agent-Definitions.md]] — Taxonomy from this history
- [[Frameworks/Framework-Comparison.md]] — Modern framework landscape
- [[Frontiers/Research-Directions.md]] — Where the field is going
- [[Domains/Coding-Agents.md]] — SWE-bench, Devin lineage
- [[Frontiers/Self-Improving-Agents.md]] — Constitutional AI → RLAIF lineage

---

*Sources: Original papers, conference proceedings (NeurIPS/ICML/ICLR/ACL/AAAI), framework release notes, GitHub trending, AI newsletters (The Batch, Import AI, TLDR AI), practitioner blogs (Lilian Weng, Andrej Karpathy, Simon Willison, Chip Huyen).*