GitHub - shareAI-lab/learn-claude-code: Bash is all you need - A nano claude code–like 「agent harness」, built from 0 to 1

9 hours ago

An agent is a trained model (neural network) that perceives, reasons, and acts, not a framework or prompt chain.
Historical examples like DeepMind's DQN, OpenAI Five, AlphaStar, and Tencent's Jueyu demonstrate agents as learned models, not scripted systems.
Modern LLM agents (Claude, GPT, Gemini) follow the same pattern: a model placed in an environment with tools.
Prompt plumbing and no-code platforms create brittle, non-generalizable systems, not true agents.
Agent development involves either training the model or building the harness (environment with tools, knowledge, context, permissions).
A harness provides the agent with tools, domain knowledge, context management, and permission boundaries for a specific domain.
Claude Code exemplifies an effective harness that trusts the model and focuses on providing tools and context without imposing rigid workflows.
Harness engineering principles are universal and applicable across domains like agriculture, healthcare, and logistics.
The minimal agent loop involves the model deciding when to use tools and when to stop, with the harness executing tool calls.
This repository teaches 12 progressive sessions on harness mechanisms, from basic loops to team coordination and worktree isolation.

Hasty Briefsbeta