A quick word on the agent harness

Two of the articles use the term agent harness, so it is worth pinning down. The harness is the software that wraps the model. The model does the reasoning and produces a tool call (a structured request to run some action). The harness is the code that actually executes that action, feeds the result back, manages the message history, and applies all the context engineering logic. When the articles talk about "managing context," they mean work the harness does, not something the model does on its own.