Protocols & Patterns

AI Harnesses

A harness is the software wrapper around a stateless LLM: state management, tool execution, guardrails, memory, and the loop that keeps the system running. This topic covers evaluation harnesses and agent harnesses, the autonomy spectrum that separates real agents from marketing, the ReAct loop architecture, and the shift from prompt engineering to writing verification loops.

In this topic

LLM harnesses, autonomous agents, and the ReAct loop architecture

Reference

What a harness is and why it matters more than the model. The two types (evaluation and agent). The autonomy spectrum: linear chains (fake agents), controlled state graphs (practical enterprise agents with LangGraph), and pure autonomous ReAct (true agency with Claude Code). The ReAct loop mechanics with an execution trace showing how observations redirect the agent's path dynamically. The modern paradigm shift from writing prompts to writing verification loops.

Last updated June 7, 2026