A Control-Theoretic Foundation for Agentic Systems

DeepMind researchers Ali Eslami and Jiangbo Yu have released a foundational paper titled 'A Control-Theoretic Foundation for Agentic Systems,' published on arXiv. The work tackles a core challenge in modern AI: how to mathematically model and analyze increasingly autonomous 'agentic' systems that can make decisions, use tools, and adapt their own goals. The researchers propose interpreting agency as hierarchical decision authority within a feedback control loop, a classic engineering concept. They introduce a unified dynamical representation that incorporates an agent's memory, learning processes, tool activation, and interaction signals into a single, analyzable closed-loop structure.

Based on this representation, the paper defines a concrete five-level hierarchy of agency. This spectrum ranges from Level 1 (simple, reactive rule-based control) up to Level 5, where an agent can synthesize entirely new control objectives and reconfigure its own decision-making architecture on the fly. The framework is presented for both nonlinear and linear systems, allowing agent behaviors to be described using standard control theory constructs like feedback gains, switching signals, and adaptation laws.

The analysis reveals that increasing an agent's autonomy introduces specific dynamical mechanisms, including time-varying adaptation, endogenous (internally-triggered) switching between strategies, decision-induced delays, and structural reconfiguration of the control pipeline itself. This perspective is significant because it provides the mathematical tools—grounded in decades of control theory—to rigorously analyze critical properties like stability, safety, and performance in AI-enabled systems. It moves the discussion of AI agents from metaphorical to mathematical, offering a formal language to reason about their behavior and potential failures.