arXiv:2609.13406 Defines Generalized Agent Iteration via Two Binary Dials

The paper models every agent as a configuration of modifiable components and every learning step as an evaluate-improve loop. Two binary conditions determine placement: whether the improvement operator resides inside the agent boundary and whether the evaluation standard is supplied externally. Systems with external improvement and external grounding map to classical GPI; systems with internal improvement and external grounding map to anchored RSI; internal improvement with internal grounding produces fully self-referential loops.

Existing architectures can now be plotted on the same plane. Standard RL agents sit in the GPI quadrant. Current open-ended agents with fixed reward models occupy the goal-drift band. No deployed system yet satisfies both internal conditions simultaneously. The framework isolates three distinct failure modes—one per missing external anchor—rather than treating RSI as a monolithic risk.

The coordinate system supplies a common language for comparing 2025-2026 agent papers without reference to marketing claims. It also identifies the minimal additional mechanism required to move any given system from one cell to another. Subsequent work can therefore target specific dial flips instead of unspecified self-improvement.

Deployment records will test the framework once agents begin logging both their improvement operators and their grounding sources at each iteration.