Text becomes action through an execution system

  • Theo starts with a simplified model of an LLM: it receives text and generates more text, including predicted tokens. Text generation alone does not read files or execute commands. 10:05 10:30
  • A harness provides the connection between model output and actions. His analogy is a chef who needs utensils and ingredients, or a programmer who needs a keyboard to act on a computer. 11:50 12:18 13:39
  • The harness describes a special response syntax. When the model uses that syntax, the harness performs the corresponding operation, such as reading or editing a file, running a command, accessing memory, or pushing to GitHub. 13:39 14:10
  • Tool access determines what the model can do in its environment. An LLM without that connection can describe an action but cannot perform it through words alone. 13:11 14:40

Minimal execution relationship

The diagram summarizes Theo's explanation of tool-call syntax and execution. 13:39 14:10

flowchart LR
    H[Harness describes tools and syntax] --> M[LLM generates a tool call]
    M --> P[Harness interprets the call]
    P --> A[Read a file, edit code, or run a command]