Harness Engineering
lastmod 2026-11-03created 2026-03-2612 edits
What is a harness?
The meaning of "harness" in AI development
A harness originally refers to the tack fitted to a horse, that is, equipment for safely controlling a horse's power and guiding it in a useful direction. As a verb, it also means "to control something powerful and put it to beneficial use."
The reason the word is used in AI development can be found right in its origin. AI agents are powerful, but left unattended they sometimes behave in unpredictable directions. A harness does not suppress that power; the term refers to the structure for steering it in the right direction while making the most of it. In other words, a harness is the entire control structure designed so that an AI agent operates in a safe and predictable way.
The role of a harness
A harness does not merely run an AI agent; it handles the following three functions in an integrated way.
- Control: a structure that keeps the agent from acting outside the permitted scope
- Monitoring: a structure that tracks and records the agent's operating state and outputs in real time
- Feedback: a feedback loop that detects errors or abnormal behavior and reflects them in the next action
The harness functions as a "safety net" and an "operational foundation" that structurally contains the unpredictable behavior and errors that arise when running agents.
See also
- Prompt engineering
- Context engineering
- Harness Engineering
- Loop Engineering
- Graph Engineering
- Data Scientist#The data scientist strikes back
- Harness - A Claude Code agent team & skill architect plugin
- Hermes Agent - A self-learning autonomous AI agent that creates and improves skills from experience
- Gstack - An open-source software factory made of AI, built and used by YC CEO Garry Tan, set up so that one person works like a team of 20.
- Korean: 컨텍스트 부패 (Context Rot)
- Prospero - A new coding harness with a lightweight design and support for remote execution
- Herdr
Favorite site
- Harness design for long-running application development | GeekNews
- [Original] What is harness engineering? 3 key elements of AI agent development
- Anthropic developed a GAN-inspired multi-agent structure to tackle two problems at once: improving frontend design quality and long-running autonomous coding
- A structure that separates the generator and the evaluator, making subjective design quality gradable against concrete criteria and solving the agent's self-evaluation bias
- A three-agent planner-generator-evaluator architecture completes full-stack applications in multi-hour autonomous coding sessions, including sprint contract negotiation and Playwright-based QA
- Moving from Opus 4.5 to Opus 4.6 made consistent coding for over 2 hours possible without sprint splitting, reducing harness complexity while maintaining performance
- Even as model performance improves, the space of interesting harness combinations does not shrink but moves, and the core task of an AI engineer is to find new combinations