---
title: Harness Engineering
type: vocabulary
url: "https://www.envisioning.com/vocab/harness-engineering"
summary: The discipline of designing and evolving an agent harness, the runtime that couples an LLM to the world through a loop, tools, context management, safety controls, orchestration, and extensions.
year: 2026
generality: 0.45
---

# Harness Engineering

The discipline of designing and evolving an agent harness, the runtime that couples an LLM to the world through a loop, tools, context management, safety controls, orchestration, and extensions.
Harness engineering is the discipline of designing and evolving an agent harness: the runtime code that couples a language model to the world through a loop, tools, context management, safety controls, orchestration, and extension surfaces. The term names the engineering practice itself, distinct from the harness as an artifact. Paul Barbaste, Tristan Darrigol, Germain Vu, and Tom Wiltberger, of Wavestone's AI Lab, gave the discipline its most detailed empirical grounding to date in "Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents," posted to arXiv (2609.00006), an expanded second edition of an April 2026 study, with the eleven-system version posted 15 July 2026. It remains a preprint, not yet peer reviewed. The same name independently appears in two other 2026 papers, "Agentic Harness Engineering" (arXiv:2604.25850) and "AI Harness Engineering: A Runtime Substrate for Foundation-Model Software Agents" (arXiv:2605.13357), suggesting the label converged into general use around the same time rather than being coined once.

The Wavestone paper performs a source-code anatomy of eleven production coding harnesses: Claude Code, Codex CLI, Gemini CLI, Mistral Vibe, OpenHands, Aider, Mini-SWE-Agent, Hermes, Pi, OpenCode, and OpenClaw, plus Databricks' Omnigent as a meta-harness contrast point, covering roughly four million lines of Python, TypeScript, and Rust. The paper defines seven canonical subsystems that recur across all eleven: the loop, the LLM-provider layer, tools, memory and context management, safety controls, orchestration of sub-agents, and extension surfaces such as hooks, skills, and the Model Context Protocol. Each subsystem is mapped from its minimal implementation (a bare while loop, a single bash tool) to its most elaborate (event-sourced replay logs, 43 typed tools loaded on demand, policy rules plus a reviewer model plus an OS sandbox).

The study reports 13 cross-cutting observations and a catalog of 29 recurring design patterns. It finds two consistent absences across all eleven systems: none imports a general-purpose agentic framework, and none retrieves code using vector embeddings, relying instead on hand-rolled loops and deterministic retrieval. A controlled longitudinal comparison of the same harnesses across one quarter shows convergence hardening into imitation, with behavioral policy migrating from prompt text into configuration files. The paper concludes that the coding harness completed a turn from tool to platform in the first half of 2026, and closes with 18 design recommendations and a 90-line minimum-viable-harness scaffold covering all seven subsystems.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/harness-engineering)