An application in Lisp you grow by talking to it

Geoffrey Huntley presents Jiti, a small Common Lisp kernel that lets you grow a running application by conversing with a language model. You request a capability, the model proposes or writes Lisp, and the kernel validates and applies that change so the application permanently acquires the new behavior until explicitly removed. The system can start from a clean slate or extend an existing image; once the model-produced definitions are accepted they become ordinary Lisp functions callable without further inference. Huntley demonstrates simple compositions (e.g., uppercase-string, reverse-string, then saving (defun shout-backwards …)) to show how an executable vocabulary accumulates through interactive use.

The architecture wires an OpenAI model, a controller, and a persistent worker: the model receives user intent, tool descriptions and observed state, then issues tool calls; the controller validates arguments and routes single actions to the worker, which evaluates forms inside a managed Lisp world and returns observations. The worker’s world adapter owns the function catalog, checkpoint/restore, durable revisions and live restarts; a reference adapter exposes named functions and data. The agent interface distinguishes develop_form (change code) from execute_form (call existing code), and caller safety checks plus goals decide acceptance. Source is on GitHub, and the larger claim is that AI-driven, in-product development eliminates traditional compile/CI loops by leveraging Lisp’s decades-old interactive facilities.