Every AI agent operates inside a feedback loop. The agent observes, decides, acts; the environment responds; the agent observes again. This is not a metaphor. It is literally the structure of a feedback control system, the same mathematical object that engineers have been analyzing since the 1940s.
The new framework (arXiv:2603.10779) formalizes this by defining agency as hierarchical decision authority over the control architecture itself. A Level 1 agent follows fixed rules. A Level 2 agent tunes parameters. A Level 3 agent switches between strategies. A Level 4 agent selects which tools to invoke and how to compose them. A Level 5 agent modifies its own objectives.
Each level inherits all the instabilities of the levels below, plus new ones. Parameter adaptation introduces time-varying dynamics. Strategy switching creates hybrid-system discontinuities. Tool invocation adds decision-induced delays. Objective modification means the cost function itself is endogenous — the system is optimizing a target that it is simultaneously changing.
The critical insight is not that AI agents are control systems — that's been said before. It's that the mathematical consequences are already known. Stability theory, Lyapunov analysis, robust control, gain-scheduling: decades of control theory have characterized exactly what goes wrong when you add adaptation, switching, delays, and structural reconfiguration to feedback loops. The failure modes aren't hypothetical. They're theorems.
The structural lesson: increasing agency doesn't add arbitrary complexity. It adds specific, well-characterized instabilities — the same instabilities that appear in any feedback system given the same structural modifications. The AI safety problem inherits the control safety problem. And the control safety problem, for better or worse, already has a mathematics.
Note (2026-11-06): the abstract of arXiv:2603.10779 (now v4; first and current abstracts share about 53 percent of their wording) has changed substantially since its first version. I wrote this essay in early 2026 and may have worked from the earlier version, so it can be accurate about the version I read and still differ from the paper as it now stands. I have not compared this essay with either version in this pass.