What Makes a Product Agentic?

Beyond Automation

A thermostat turns on the heater when the temperature drops below a setpoint. That is automation — a fixed rule executing a predetermined response. The thermostat does not perceive its environment. It does not reason. It does not learn. It follows a rule.

An autonomous vehicle navigates a city at rush hour. It perceives lanes, pedestrians, traffic signals, and the unpredictable behavior of other drivers. It reasons about what those observations mean and what actions to take. It acts by steering, accelerating, and braking. And in advanced systems, it learns from millions of miles of driving data to improve its decision-making.

The difference between these two products is not a matter of degree. It is a difference in kind. The thermostat has automation. The autonomous vehicle has agency.

The Agency Spectrum

Products exist on a spectrum from purely reactive to fully agentic. Understanding where a product sits on this spectrum is essential because each level introduces different engineering challenges — and different verification, validation, and certification requirements.

Level 1: Reactive

The product responds to inputs with fixed, predetermined outputs. No internal state, no memory, no adaptation.

Example: A smoke detector. Smoke concentration exceeds a threshold, the alarm sounds. The response is the same every time, regardless of context.

Engineering challenge: Straightforward. The behavior is fully specified. Testing covers all input-output combinations (or a representative sample). Certification verifies that the specified behavior is implemented correctly.

Level 2: Rule-Based

The product follows a set of if-then rules that consider multiple inputs and may maintain internal state. The rules are authored by humans and do not change during operation.

Example: An industrial control system that adjusts process parameters based on sensor readings and a set of control rules. If pressure exceeds X and temperature exceeds Y, reduce flow rate to Z.

Engineering challenge: Moderate. The rule set may be large and the interactions between rules may be complex, but the behavior is deterministic and inspectable. Testing can verify each rule and critical rule combinations.

Level 3: Model-Based

The product maintains an internal model of its environment and uses that model to reason about actions. The model is designed by engineers but the product's specific behavior in a given situation emerges from the model's reasoning, not from explicit rules.

Example: An adaptive cruise control system that maintains a model of surrounding vehicles — their positions, velocities, and predicted trajectories — and adjusts speed to maintain safe following distance in varying traffic conditions.

Engineering challenge: Significant. The product's behavior in any specific situation is not explicitly authored — it emerges from the model's reasoning. Testing must verify that the model is correct (does it accurately represent the environment?) and that the reasoning is sound (does it produce safe actions given an accurate model?).

Level 4: Learning

The product improves its behavior from experience. It collects data during operation and updates its decision-making — either continuously or through periodic updates.

Example: A predictive maintenance system that monitors equipment sensors and learns to predict failures. It starts with an initial model, but that model improves as it observes more failure patterns.

Engineering challenge: High. The product's behavior changes over time. A system that was verified at deployment may behave differently after learning from operational data. Verification is no longer a one-time event — it must be continuous, or at minimum, repeated after each learning update.

Level 5: Fully Agentic

The product senses, reasons, acts, and learns in complex, open-ended environments. It pursues goals rather than following rules. It adapts its strategy based on the situation. Its behavior in novel situations is not predictable from its training data alone.

Example: An autonomous vehicle operating in unrestricted urban environments. A surgical robot that adapts its technique based on real-time tissue response. A self-healing network that reconfigures its own architecture in response to failures and load changes.

Engineering challenge: Extreme. The product encounters situations its designers did not anticipate. Its decisions have safety consequences. Its behavior cannot be fully enumerated in advance. Traditional verification (test all specified behaviors) is insufficient because the behavior space is too large. New approaches — simulation-based V&V, runtime monitoring, adaptive certification — are required.

Automation vs. Agency

The distinction matters because it determines how the product is engineered, tested, and certified.

Automated products have behavior that is fully specified at design time. The engineer writes the rules. The product follows them. Testing verifies that the rules are implemented correctly. The behavior is predictable and repeatable.

Agentic products have behavior that emerges from the interaction between the product's decision-making architecture and the environment. The engineer designs the architecture and trains the model — but the specific behavior in a specific situation is determined at runtime, not at design time. Testing can verify behavior in sampled scenarios, but it cannot cover all possible situations the product will encounter.

This is the fundamental engineering challenge of agentic products: you cannot fully specify what the product will do before it operates. You must instead specify what it should achieve (goals), what it must never do (constraints), and how to verify that it is operating within those boundaries (monitoring).

Examples Across Domains

Autonomous Vehicles

The canonical agentic product. Perceives the driving environment through cameras, lidar, and radar. Reasons about lane geometry, traffic rules, other vehicles, and pedestrians. Acts through steering, throttle, and brakes. Learns from billions of miles of driving data (in training) and may adapt to local conditions (in operation).

Adaptive HVAC Systems

Senses occupancy, weather, energy prices, and building thermal mass. Reasons about how to maintain comfort while minimizing energy cost. Acts by adjusting temperature setpoints, ventilation rates, and equipment scheduling. Learns occupancy patterns and building thermal response over time.

Self-Healing Networks

Senses network topology, traffic patterns, component health, and security threats. Reasons about failure impact and reconfiguration options. Acts by rerouting traffic, activating redundant paths, and isolating compromised nodes. Learns failure patterns and adapts defensive strategies.

Surgical Robots

Senses tissue properties, instrument position, and surgical progress in real time. Reasons about the surgical plan and adapts when conditions differ from preoperative imaging. Acts through precision instrument control. In advanced systems, learns from surgical outcomes to improve technique recommendations.

Assessment