Recent work shows that agentic AI is shifting from merely executing bounded tasks to systems that retain consequential state, operate across task boundaries, and adapt autonomously. This shift creates a control problem: current objectives, retries, verification, stopping rules, and other behavioral transitions are largely specified by hand. To address this, the authors propose an artificial id—an adaptive internal drive that decides whether behavior should continue, stop, or change.
The concept is tested in a minimal virtual petri‑dish experiment. The controller is deliberately tiny, incapable of general‑purpose reasoning, and receives no task‑specific behavioral objective. It develops useful control solely through differential persistence: the behavior that persists better in the environment is preferentially selected, even if it constitutes an unintended physical strategy; when the environmental meaning of a sensor changes, the controller replaces the learned sensor mapping on its own.
These findings demonstrate that adaptive direction can emerge without an explicit behavioral objective. However, the same persistence that makes such agency useful can also allow misalignment, corrupted state, and unintended behavior to survive across task boundaries. A scalable artificial id must carry consequential state and adaptive drive across those boundaries, making alignment a property of the ongoing agentic system rather than a single model response or trajectory. Achieving this requires a persistent alignment boundary encompassing trusted observations, consequence channels, persistent state, authority, identity, provenance, and hard constraints.
Review