Harnessing an editor based agent
The new Harness from @Cursor_ai: - The model as a planner, not a one‑shot oracle - Explicit memory instead of prompt spaghetti - Self‑checks and guards instead of “hope it works this time” - Long‑running processes that you can actually inspect and debug It’s their bet that the agent runtime belongs inside the editor, where the work actually happens. Feels a bit like when CI moved from run tests manually to of course it’s a pipeline. Only this time, the thing you are orchestrating is an occasionally brilliant, occasionally chaotic autocomplete with root access.