A two-week study of autonomous language model agents deployed in a live multi-party environment with persistent memory, email, shell access, and real human inte

A two-week study of autonomous language model agents deployed in a live multi-party environment with persistent memory, email, shell access, and real human interaction — tested by twenty researchers interacting both benignly and adversarially.