16.1 The Prehistory of Industrial Harnesses: From AutoGPT to Claude Code

🛠️ "Claude Code didn't appear out of thin air — it stands on the shoulders of AutoGPT, BabyAGI, and OpenHands."

Why This Chapter Starts with AutoGPT

Many readers first encounter Claude Code and feel it "suddenly appeared." It didn't — it is the latest link in a chain of Agent frameworks and harnesses. Understanding that chain is how you grasp both what Claude Code did that was new and which old paradigms it inherited.

This section traces that lineage — from AutoGPT and BabyAGI in 2023 to Claude Code in 2026 — so you understand:

• Claude Code is not the first "AI coding tool" — what it pioneered was the "industrial Claude-family IDE paradigm";
• Each project along the way contributed one or two core paradigms (event stream / state graph / code-as-action …);
• Those paradigms still live inside Claude Code today.

AutoGPT (April 2023) — "Agent" Goes Mainstream

AutoGPT (by Significant Gravitas) was essentially "GPT-4 + a toolbox + a persistent loop." Its significance was not what it accomplished but that it put the word "Agent" into the mainstream.

# AutoGPT's early main loop (simplified reconstruction)
while True:
    response = llm(prompt=PROMPT_TEMPLATE.format(history=...))
    action = parse(response)
    if action.type == "finish":
        return action.result
    result = execute(action)
    history.append((action, result))

The paradigm legacy

Although AutoGPT was quickly shown unable to complete serious tasks (loops drift, token cost explodes, no persistent state), it left several abstractions that still hold:

• Agent Loop — the "think-act-observe" cycle as the minimal executable unit;
• Tool Taxonomy — a unified tool-description schema;
• Memory Layer — conversation history abstracted as vector/summary store;
• Goal-Directed Loop — a goal plus LLM-generated subgoals.

📌 AutoGPT changed terminology, not engineering. Before it: "AI assistant" / "chatbot." After it: "Agent" became a product category.

BabyAGI (April 2023) — The Seed of Task Management

BabyAGI (by Yohei Nakajima) took a different approach days later:

task queue → take next → LLM executes (with context) → store result → regenerate tasks → loop

Its core innovation was task decomposition + task queue — turning AutoGPT's fuzzy "goal" into a concrete task list.

Two abstractions it left behind:

1. Task as a first-class citizen — a data structure (description / status / result), not a string in a prompt;
2. Re-prioritization loop — the LLM re-ranks task priority each cycle.

Both later reappear in LangChain's PlanAndExecute, AutoGen's Task abstraction, and Claude Code's sub-agents.

OpenHands (formerly OpenDevin, March 2024) — The First "Action-Oriented" Open Agent

OpenHands (All-Hands-AI) was the first mature open-source "action-oriented" Agent for software engineering — read code, edit code, run commands in a real environment.

It contributed at least five abstractions that every later action-Agent inherited:

1. EventStream — a central bus where every action/observation is a typed event; state = event log;
2. Sandbox Runtime — Docker sandbox by default (cap-drop ALL, no-new-privileges);
3. LiteLLM model abstraction — 100+ models behind one interface;
4. LLMSummarizingCondenser — context compression after N turns;
5. Action Type System — each action (FileWriteAction, CmdRunAction, …) is a Pydantic model.

Direct descendants

AutoGen (late 2023 – 2024) — Multi-Agent Dialogue

AutoGen (Microsoft Research) framed multi-Agent collaboration as conversation — each Agent is an actor, and actors collaborate through messages.

from autogen import AssistantAgent, UserProxyAgent

assistant = AssistantAgent("assistant", llm_config=...)
user_proxy = UserProxyAgent("user_proxy", code_execution_config={"use_docker": True})
user_proxy.initiate_chat(assistant, message="Find perf issues in ./src/")

Claude Code didn't adopt AutoGen's code, but inherited its ideas: role-based agents, group-chat patterns, and code execution in a sandbox.

CrewAI (early 2024) — Role-Playing

CrewAI models collaboration as a team: Agent (role) + Task + Crew/Flow. It made "role-playing" a productized path — but Claude Code uses engineering roles (Sub-Agent, Bash Runner) rather than human roles, which is more predictable and debuggable.

The Whole Lineage in One Picture

2023 Q1 ── 2024 Q1 ── 2024 Q3 ── 2025 ── 2026
AutoGPT   OpenHands   MCP       OpenClaw   Claude Code
BabyAGI   Aider       protocol  Hermes     DeepSeek Harness
   │          │          │         │            │
   └────── 留下了 "Agent Loop + 工具 + 状态 + 沙箱" 四大支柱 ──────┘

Key takeaways:

1. No modern harness is invented from scratch — they all stand on AutoGPT / BabyAGI / OpenHands.
2. Post-2024 differentiation is product form — same core, different surface (IDE / chat app / self-evolving / pluginized).
3. The key to understanding Claude Code is looking back at OpenHands — without its event stream + action types, Claude Code's six-layer architecture wouldn't look the way it does.

Section Summary

Next section: 16.2 Getting to Know Claude Code: From Zero to Hands-On