A model on its own is a brilliant advisor locked in a room. The harness is the assistant standing outside the door: it carries your request in, reads what the advisor wants done next, does it (or asks you first), carries the result back in, and repeats until the job is finished.
Concretely, the harness:
- sends your prompt and the current context to the model,
- notices when the model says “I need to read this file” or “run this command”,
- asks you for permission when the action could change something, as far as the permission mode requires,
- performs the action using a tool and gives the result back to the model,
- loops until the model says it is done, then shows you the outcome.
This loop, gather context then act then check, is what the word agentic means. The agent loop note shows it step by step, with Anthropic’s diagram. Claude Code is a harness. The chat on claude.ai is a much thinner harness with fewer tools, and the programs it runs stay in a sandbox in the cloud. See Chat, Work and Code.
Example
You ask for “a summary of every invoice in this folder as a table”. The harness lets the model list the folder, read each file, and write a new file with the table. You approve the first write. The model never touched your disk; the harness did, on its instructions and with your consent.
Why it matters for you
The harness is where safety and control live. It decides which tools exist, what needs your approval and what runs automatically. When you understand that, “AI on my computer” stops sounding like magic or a threat: it is a program that does exactly the actions you allow.
Related
- Agent loop - the loop the harness runs, step by step
- Tool - the actions a harness can perform
- Permission mode - the dial that decides when it stops to ask
- Session - the harness saves the conversation and snapshots files before it changes them
- Slash command - how you give the harness itself an instruction
- Agent - model plus harness plus tools, working in a loop
- Mental models - AI = model + harness + tools