Every AI agent is just patches to one JSON loop, argues engineer
Harness Engineering 101: The LLM is the brain, the harness is the body
A 16-chapter book by Isuru Wijesiri builds an AI agent from first principles, starting from the fact that an LLM API is stateless: each turn you resend the whole conversation as a JSON array. The "harness" is the program that builds and protects that array, and every pattern the field calls an agent is a patch to that loop, each added because something concrete broke. A ~300-line Python toy agent grows along the way.
Everything the field calls "agents" is a set of patches to that one loop, and each patch exists because something concrete broke.