Chapter 12 · Tools, agents and harnesses
The agent harness
The harness is the software around a model: it runs the loop, offers tools, manages context, permissions and memory.
Preparing the agent harness
Chapter 12 · Tools, agents and harnesses
The harness is the software around a model: it runs the loop, offers tools, manages context, permissions and memory.
The model in the middle can only read and write text. The harness around it passes your request in, turns tool requests into real actions, feeds results back, and checks with you before anything risky.
On its own, a language model cannot open a file, search the web or click a button. It can only write.
Task: 'Add up my three invoices and save a summary.' Switch parts off and see how the same model copes.
Finished safely
Who actually opens a file when an agent 'reads' it?
What happens without the agent loop?
Why do harnesses ask permission?