What is an AI harness

A harness is the program that surrounds an AI model and lets it do things: it gives it tools, memory, file access and permissions, and it manages the cycle of asking, acting and checking the result. It's often summed up like this: an agent is a model plus a harness. The model thinks and writes; the harness turns that into actions.

A model on its own only produces text and remembers nothing from one conversation to the next. The harness is what lets it read a file, run a command, search the internet or save what it has done, and also what sets limits so it doesn't do what it shouldn't. Claude Code, OpenClaw and DeepSeek Harness are examples of harnesses for agents. What's confusing is that the word is used rather loosely: on this blog I also call harnesses, in a broad sense, the programs that load and run local models, like Ollama or llama.cpp.

I work every day with several harnesses, in the cloud and locally, spread across my in-house servers and my VPS. How it went when I brought a new one into my real work I explain in "Ten days working with DeepSeek Harness"; the ones I use for local AI, in "Ollama, LM Studio, llama.cpp and Unsloth: which one I use for what"; and the map of what runs on each machine and why, in "What runs on each of my machines".