A Control-Theoretic Foundation for Agentic Systems

This paper establishes a control-theoretic framework that formalizes agentic systems as hierarchical decision authorities within feedback loops, providing a unified dynamical representation to analyze the stability, safety, and performance of AI agents that adapt controllers, strategies, and objectives in real-time.

The Gist

Imagine you are driving a car. In a traditional car, the steering wheel, pedals, and brakes are connected directly to the wheels. If you turn the wheel, the car turns. The "rules" of how the car responds are fixed by the engineer who built it.

Now, imagine a new kind of car where an AI Agent is sitting in the driver's seat. But this isn't just a GPS telling you where to go. This AI is a super-driver that can do much more than just steer.

This paper, titled "A Control-Theoretic Foundation for Agentic Systems," is essentially a rulebook for understanding what happens when we let this super-driver take control of the car's very DNA. It asks: What happens to the car's stability when the driver can change the engine, swap the tires, rewrite the map, and even decide where they want to go, all while driving?

Here is the breakdown of the paper's ideas using simple analogies:

1. The Core Idea: The "Control Stack"

Think of the car's control system as a layered cake:

• Bottom Layer: The engine and wheels (the physical car).
• Middle Layer: The steering and brakes (the controller).
• Top Layer: The destination and the rules of the road (the goals).

In old-school AI, the AI was just a passenger giving a voice command like "Turn left." In Agentic Systems, the AI is the driver who can reach down and tweak the engine, swap the steering mechanism, or decide to drive to the beach instead of work.

The authors created a mathematical framework to describe this. They say that as the AI gets more "agency" (power), it climbs up this cake, taking control of higher and higher layers.

2. The 5 Levels of "Super-Driver" Power

The paper defines five levels of how much control the AI has. Let's use a Video Game Character analogy to explain them:

• Level 1: The NPC (Non-Player Character)

  • The Analogy: The AI is like a basic video game character programmed to "Walk forward if no wall." It has no brain. It just follows a strict script.
  • The Tech: The AI just reacts to what it sees. No learning, no changing goals.

• Level 2: The Tuner

  • The Analogy: The AI is now a mechanic who can tighten the bolts. It can't change the car model, but it can make the suspension stiffer or softer depending on the road.
  • The Tech: The AI can adjust the numbers (parameters) inside the controller to make it work better, but the overall design stays the same.

• Level 3: The Strategist

  • The Analogy: The AI is now a tactical commander. It has a menu of pre-made cars (a race car, a tank, a truck). If it sees a mountain, it swaps the race car for the truck. If it sees a track, it swaps to the race car.
  • The Tech: The AI can choose between different pre-set strategies or goals (e.g., "Drive fast" vs. "Drive safely").

• Level 4: The Architect

  • The Analogy: The AI is now a custom builder. It doesn't just pick a car; it builds a new one on the fly. It might decide, "I need a car with a radar, a turbo, and a special brake system," and it assembles these parts together in a new way.
  • The Tech: The AI can rewire the system, connecting different tools and modules in new combinations to solve a problem.

• Level 5: The Visionary

  • The Analogy: The AI is now the Game Designer. It doesn't just play the game; it decides what the game is about. It might say, "Actually, the goal isn't to reach the finish line; the goal is to collect as many coins as possible while avoiding the red zone." It invents new rules and new objectives on the spot.
  • The Tech: The AI can generate entirely new goals and system structures, as long as they follow safety rules (governance).

3. The Danger Zone: Why This Matters

The paper's most important warning is about Stability.

Imagine you are driving a car.

• If the AI changes the suspension (Level 2) too fast, the car might start shaking violently.
• If the AI keeps switching between a race car and a truck (Level 3) every second, the car might lose control and spin out, even if both the race car and the truck are safe on their own.
• If the AI keeps rebuilding the engine while driving (Level 4), the car might fall apart.

The authors show that more power for the AI doesn't automatically mean a better car. In fact, if the AI changes its mind too quickly or restructures the car too often, the whole system can become unstable and crash.

4. The "Time Travel" Problem

The paper also points out that thinking takes time.

• If the AI has to stop, look up a map, call a friend, and then decide where to go, there is a delay.
• In control theory, delays are dangerous. If you turn the steering wheel, but the car waits 2 seconds to respond, you might over-correct and crash.
• As the AI gets smarter (Levels 3–5), it does more thinking, which creates more delays, which makes the car harder to control.

The Big Takeaway

This paper is a mathematical safety manual for the future of AI.

It tells engineers and scientists: "Don't just throw an AI into a control system and hope for the best. You need to understand what kind of power you are giving it.

• If you give it Level 2 power, watch out for shaking (instability from rapid changes).
• If you give it Level 3 power, watch out for spinning out (instability from switching too fast).
• If you give it Level 5 power, you need a very strict rulebook (governance) to make sure it doesn't invent a goal that gets everyone hurt."

In short, the paper provides the tools to build AI drivers that are not only smart but also safe and predictable, ensuring that as they gain more power, the car doesn't crash.

Technical Summary

Here is a detailed technical summary of the paper "A Control-Theoretic Foundation for Agentic Systems" by Ali Eslami and Jiangbo Yu.

1. Problem Statement

Modern AI systems are increasingly deployed within feedback control loops, transitioning from static design-time assistants to runtime decision-makers. Unlike traditional controllers, agentic AI systems possess the authority to:

• Adapt controller parameters online.
• Select among different control strategies or policies.
• Invoke external computational tools (e.g., planners, estimators).
• Reconfigure the decision architecture (workflows).
• Modify control objectives based on context or interaction.

The Core Challenge: Traditional control theory offers distinct frameworks for fixed controllers, adaptive systems, switched systems, and hybrid systems. However, agentic systems often combine these mechanisms simultaneously within a single loop. The decision-making process itself becomes part of the feedback loop, rendering existing models (which assume fixed structures or simple switching) insufficient. There is a lack of a unified mathematical framework to define "agency," analyze how increasing decision authority affects closed-loop dynamics, and assess stability, safety, and performance.

2. Methodology

The authors propose a unified control-theoretic framework that interprets agency as hierarchical decision authority over the control architecture.

A. Unified Dynamical Representation

The paper formulates a general nonlinear model for an agentic controller embedded in a physical plant:
$$\dot{x}(t) = f(x(t), u(t), w(t))$$
$$y(t) = h(x(t), w(t))$$

The controller acts on an Information Set $I(t) = \{y(t), m(t), z(t), r(t)\}$, comprising:

• $y(t)$: Observations.
• $m(t)$: Internal memory (beliefs, context, state estimates).
• $z(t)$: Outputs from external tools.
• $r(t)$: Interaction signals (human instructions, multi-agent communication).

The control law is defined as:
$$u(t) = \pi_{\alpha(t)}(I(t); \theta(t), \zeta(t))$$
Where:

• $\theta(t)$: Adaptable parameters (learning).
• $\zeta(t)$: Goal descriptor (defining the cost function).
• $\alpha(t)$: Active controller architecture/policy family.
• $\sigma(t), c(t)$: Tool activation and workflow composition variables.

B. The Five-Level Hierarchy of Agency

The authors define a hierarchy based on the degree of authority the AI holds over the variables $\{\theta, g, \zeta, \sigma, c, \alpha\}$:

1. Level 1 (Reactive Rule-Based): Fixed rules. The AI executes predefined if-then logic. No adaptation, learning, or goal modification.
2. Level 2 (Adaptive Agency): Fixed structure, but parameters ($\theta$) and memory ($m$) adapt online. The AI tunes gains or weights within a fixed template.
3. Level 3 (Strategic Selection): The AI selects among predefined sets of controllers ($\alpha$), goals ($\zeta$), and tools ($\sigma$) based on context.
4. Level 4 (Structural Agency): The AI reconfigures the decision architecture itself. It composes workflows, sequences tools, and alters the structural interconnection of modules (e.g., switching between direct feedback and estimator-based control).
5. Level 5 (Generative Agency): The AI synthesizes new goals, workflows, and architectures online, subject to governance constraints. It can redefine the problem to be solved.

C. Linear Specialization

To provide concrete interpretations, the framework is specialized to Linear Time-Invariant (LTI) systems. In this setting:

• Agency maps to modifying feedback gains ($K$), switching between quadratic cost matrices ($Q, R$), or reconfiguring state-space interconnections.
• Level 5 is represented as the constrained synthesis of admissible linear objectives and architectures.

3. Key Contributions

1. Unified Framework: A single dynamical representation that integrates memory, learning, tool use, interaction, and goal formulation into a closed-loop structure.
2. Formal Hierarchy of Agency: A rigorous 5-level classification that quantifies AI authority from reactive rules to generative synthesis, linking abstract AI capabilities to specific control-theoretic constructs.
3. Linear Interpretation: A specialized linear formulation that maps agency levels to familiar objects: fixed gains, adaptive gains, switched systems, and modular interconnections.
4. Stability Analysis: Identification of specific dynamical mechanisms introduced by increasing agency:
   • Level 2: Time-varying dynamics (due to parameter adaptation).
   • Level 3: Endogenous switching (due to strategy selection).
   • Level 3-5: Decision-induced delays (due to reasoning/tool latency).
   • Level 4: Hybrid dynamics (due to structural reconfiguration).
   • Level 5: Objective-induced policy variation.

4. Results and Simulations

The paper validates the theoretical framework through three simulation examples demonstrating how increased agency can lead to instability even if individual components are stable:

• Level 2 (Adaptation): A spring-mass-damper system with an adaptive proportional gain.
  • Result: Slow adaptation rates maintain stability, while fast adaptation rates destabilize the time-varying closed-loop system.
• Level 3 (Strategic Switching): A discrete-time system switching between two stabilizing objectives (state regulation vs. output tracking).
  • Result: While both controllers are individually stable, rapid switching (dwell time $d=1$) creates a switched system with a spectral radius $>1$, causing exponential instability.
• Level 4 (Structural Reconfiguration): A system switching between direct state feedback and an estimator-augmented pipeline.
  • Result: The estimator pipeline introduces internal states. Rapid switching between architectures leads to instability due to the interaction of the different internal dynamics, highlighting that structural changes alter the system order and stability margins.

5. Significance and Implications

• Bridging AI and Control: The paper provides a mathematical language to discuss "agentic" behaviors using standard control theory concepts (stability, switching, hybrid systems), moving beyond application-specific case studies.
• Safety and Verification: It highlights that increasing AI authority fundamentally changes the type of dynamical system being analyzed. Safety guarantees for fixed controllers do not automatically transfer to agentic systems.
• Design Guidelines: The framework suggests that managing agency requires specific control-theoretic constraints, such as:
  • Limits on adaptation rates (for Level 2).
  • Dwell-time constraints on switching (for Level 3).
  • Delay-aware analysis for reasoning steps (Level 3-5).
  • Constraints on architectural synthesis (Level 4-5).
• Future Directions: The authors identify the need for formal stability guarantees for higher agency levels, extensions to multi-agent settings, and methods to handle semantic ambiguity in human-AI interaction.

In conclusion, this work establishes that agency is a form of decision authority that, when embedded in feedback loops, transforms the system dynamics. Understanding these transformations is critical for ensuring the safety and reliability of next-generation AI-enabled cyber-physical systems.