Agent = model + harness

On its own, a model cannot do anything in another environment (e.g. your computer). The harness runs on your computer, keeps the context, sends it to whichever model you pick, takes the actions the model asks for and appends action outcomes to the context. This plays five rounds of that loop, starting from nothing but a system prompt.

Harness loop A user sends prompts into a harness. The harness appends each prompt to a context that starts with only a system prompt, sends the full context to the chosen model in the cloud, runs the tools the model calls (bash, edit, read, write) against files, and appends each tool result to the context before the reply goes back to the user. Harness runs the loop on your computer e.g. Claude Code OpenAI Codex ChatGPT Gemini CLI Cursor Context system prompt time model full context User prompt reply Harness model Local model GPT Gemini Claude bash edit read write Your Computer PDF.pdf .xlsx .ts request response or result Seref Arikan, PhD linkedin.com/in/seref-arikan/

Mode
Speed