So a "harness" is really a context orchestration tool. Every request body it builds typically has the same three parts:

SemiAnalysis Twitter · SemiAnalysis (@SemiAnalysis_) · 2026-07-03

SemiAnalysis breaks down the three universal components of every LLM harness request—system prompt, tool definitions, and message history—and explains that harnesses like Claude Code and Codex differ primarily in how they manage and compress this context.

Open original ↗

Appears in

SemiAnalysis Demystifies Agentic Coding Harness Architecture: Model vs. Orchestration

Extraction

Topics: llm-harnessescontext-managementtool-useagentic-coding

Claims

A harness is fundamentally a context orchestration tool, not a magic black box.
Every harness request contains the same three parts: a system prompt, tool definitions (as JSON schemas), and a chronological message history.
The key differentiator between competing harnesses like Claude Code, Codex, and OpenCode is their context management strategy, not just their UI.

Key quotes

a 'harness' is really a context orchestration tool.

Since the model is stateless, the harness reassembles and resends all of these parts every turn. What differs between Claude Code, Codex, and OpenCode is not only the TUI/UX features, but how this context is managed.