We spent months building AI agents. Then we deleted them (runnit.io)

[dead]

You missed a key point here: context bias. It's always better to have conpletely separate agents per task if you want "fresh eyes" on any given task.