← Back to Table of Contents

Chapter 2 — History & Evolution

“The best way to predict the future is to invent it.” — Alan Kay

How We Got Here

The idea of machines that reason and act isn’t new. What’s new is that LLMs finally made it work. Here’s the journey from early AI to today’s autonomous agents.

Timeline of Agent Evolution
1966
ELIZA
Joseph Weizenbaum's pattern-matching chatbot. No reasoning — just regex. But it showed people want to interact with machines conversationally.
1980s
Expert Systems
Hand-coded if/then rules (MYCIN, DENDRAL). Agents before LLMs — brittle, expensive to maintain, couldn't generalize.
1996
BDI Architecture
Belief-Desire-Intention model formalized how agents reason. Influenced robotics and game AI for decades.
2017
Transformer Architecture
"Attention Is All You Need" — the foundation for every modern LLM and, by extension, every modern agent.
2022 — Oct
ReAct Paper
Yao et al. showed LLMs can interleave reasoning (thinking) with acting (tool use). The birth of modern agents.
2023 — Feb
Toolformer
Meta showed LLMs can learn to use tools autonomously — deciding when and how to call APIs from training data.
2023 — Mar
GPT-4 + Function Calling
OpenAI released GPT-4 with native function calling. This made building agents accessible to any developer.
2023 — Apr
AutoGPT / BabyAGI
Open-source fully autonomous agents. They showed the dream (and the limitations) of self-directed AI.
2023 — May
Voyager
NVIDIA's agent that wrote its own code to explore Minecraft — learning new skills by writing and saving programs.
2023 — Oct
Generative Agents (Stanford)
25 AI agents living in a simulated town, forming relationships, planning parties. Memory + reflection = emergent behavior.
2024
Framework Explosion
LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Anthropic's tool use — agents went from research to production.
2025
MCP & Computer Use
Model Context Protocol standardized tool integration. Anthropic's computer use let agents control desktops. Agents became mainstream.
2026
Agentic Enterprise
Agents in production at scale — customer support, coding, research, operations. The "agent developer" role emerges.

The Key Papers

If you want to go deep, these are the foundational papers that shaped modern agents:

Paper Year Key Contribution
ReAct: Synergizing Reasoning and Acting 2022 Interleaving thought + action traces — the core agent pattern
Toolformer 2023 LLMs learning to use tools from self-supervised data
Reflexion 2023 Agents that learn from their own mistakes through self-reflection
Voyager 2023 Lifelong learning agent that writes and stores its own code
Generative Agents 2023 Memory + reflection → emergent social behavior
Chain-of-Thought Prompting 2022 Step-by-step reasoning dramatically improves LLM performance
Tree of Thoughts 2023 Exploring multiple reasoning paths, not just one chain

The Three Breakthroughs That Enabled Modern Agents

Why Agents Work Now
🧠
Capable LLMs
GPT-4, Claude, Gemini — models smart enough to reason over multi-step problems and follow complex instructions
🔧
Function Calling
Native support for structured tool invocation — the LLM can reliably output JSON to call functions
💾
Long Context + RAG
128K–1M token windows + vector search = agents can reason over large knowledge bases

Lessons from the Early Autonomous Agent Era

AutoGPT (2023) was one of the fastest-growing GitHub repos in history. But it also taught us hard lessons:

  1. Unbounded autonomy fails — agents would loop forever, spending money on API calls without making progress
  2. Planning > execution — agents good at individual steps but bad at long-horizon planning
  3. Hallucination is deadly for agents — a chatbot hallucination is annoying; an agent hallucination triggers real-world actions
  4. Human-in-the-loop is essential — the best agent systems keep humans at critical checkpoints

These lessons shaped the move from “fully autonomous” to “agentic workflows” — controlled autonomy with human oversight.

What Changed: The Modern Agent Paradigm

2023 vs. 2026: How Agent Design Evolved
Early 2023: The Dream
  • Fully autonomous — "set and forget"
  • One giant agent does everything
  • Free-form planning with no guardrails
  • LLM decides everything
  • Cool demos, unreliable in practice
2026: The Reality
  • Controlled autonomy with human checkpoints
  • Specialized agents collaborating
  • Structured workflows + constrained actions
  • LLM handles reasoning; code handles control flow
  • Production-grade with observability

What’s Next

With this historical context, let’s zoom into the architecture of a modern agent — what components it’s made of and how they fit together.

Next: Chapter 3 — Anatomy of an Agent →

← Previous: Chapter 1 — What Are Agents? · Next: Chapter 3 — Anatomy of an Agent →

Last updated: April 2026