Max Headroom 
We tend to assume that if every part of a system behaves correctly, the system itself will behave correctly. That assumption is deeply embedded in how we design, test, and operate software. If a service returns valid responses, if dependencies are reachable, and if constraints are satisfied, then the system is considered healthy. Even in distributed systems, where failure modes are more complex, correctness is still tied to the behavior of individual components. In modern AI systems, particularly those combining retrieval, reasoning, and tool invocation, this assumption is increasingly stressed under continuous operation.
This model works because most systems are built around discrete operations. A request arrives, the system processes it, and a result is returned. Each interaction is bounded, and correctness can be evaluated locally. But that assumption begins to break down in systems that operate continuously. In these systems, this behavior is not the result of a single request. It emerges from a sequence of decisions that unfold over time. Each decision may be reasonable in isolation. The system may satisfy every local condition we know how to measure. And yet, when viewed as a whole, the outcome can be wrong.
One way to think about this is as a form of behavioral drift systems that remain operational but gradually diverge from their intended trajectory. Nothing crashes. No alerts fire. The system continues to function. And still, something has gone off course.
The composability problem
The root of the issue is not that components are failing. It is that correctness no longer composes cleanly. In traditional systems, we rely on a simple intuition: If each part is correct, then the system composed of those parts will also be correct. This intuition holds when interactions are limited and well-defined.
In autonomous systems, that intuition becomes unreliable. Consider a system that retrieves information, reasons over it, and takes action. Each step in that process can be implemented correctly. Retrieval returns relevant data. The reasoning step produces plausible conclusions. The action is executed successfully. But correctness at each step does not guarantee correctness of the sequence.
The system might retrieve information that is contextually valid but incomplete or misaligned with the current task. The reasoning step might interpret it in a way that is locally consistent but globally misleading. The action might reinforce that interpretation by feeding it back into the system’s context. Each step is valid. The trajectory is not. This is what behavioral drift looks like in practice: locally correct decisions producing globally misaligned outcomes.
In these systems, correctness is no longer a property of individual steps. It is a property of how those steps interact over time. This breakdown is subtle but fundamental. It means that testing individual components, even exhaustively, does not guarantee that the system will behave correctly when those components are composed into a continuously operating whole.
Behavior emerges over time
To understand why this happens, it helps to look at where behavior actually comes from. In many modern AI systems, behavior is not encoded directly in a single component. It emerges from interaction:
- Models generate outputs based on context
- Retrieval systems shape that context
- Planners sequence actions based on those outputs
- Execution layers apply those actions to external systems
- Feedback loops update the system’s state
Each of these elements operates with partial information. Each contributes to the next state of the system. The system evolves as these interactions accumulate. This pattern is especially visible in LLM-based and agentic AI systems, where context assembly, reasoning, and action selection are dynamically coupled. Under these conditions, behavior is dynamic and path dependent. Small differences early in a sequence can lead to large differences later on. A slightly suboptimal decision, repeated or combined with others, can push the system further away from its intended trajectory.
This is why behavior cannot be fully specified ahead of time. It is not simply implemented; it is produced. And because it is produced over time, it can also drift over time.
Observability without alignment
Modern observability systems are very good at telling us what a system is doing. We can measure latency, throughput, and resource utilization. We can trace requests across services. We can inspect logs, metrics, and traces in near real time. In many cases, we can reconstruct exactly how a particular outcome was produced. These signals are essential. They allow us to detect failures that disrupt execution. But they are tied to a particular model of correctness. They assume that if execution proceeds without errors and if performance remains within acceptable bounds, then the system is behaving as expected.
In systems exhibiting behavioral drift, that assumption no longer holds. A system can process requests efficiently while producing outputs that are progressively less aligned with its intended purpose. It can meet all its service-level objectives while still moving in the wrong direction. Observability captures activity. It does not capture alignment.
This distinction becomes more important as systems become more autonomous. In AI-driven systems, particularly those operating as long-lived agents, this gap between activity and alignment becomes operationally significant. The question is no longer just whether the system is working. It is whether it is still doing the right thing. This gap between activity and alignment is where many modern systems begin to fail without appearing to fail.
The limits of step-level validation
A natural response to this problem is to add more validation. We can introduce checks at each stage:
- Validate retrieved data.
- Apply policy checks to model outputs.
- Enforce constraints before executing actions.
These mechanisms improve local correctness. They reduce the likelihood of obviously incorrect decisions. But they operate at the level of individual steps.
They answer questions like:
- Is this output acceptable?
- Is this action allowed?
- Does this input meet requirements?
They do not answer:
- Does this sequence of decisions still make sense as a whole?
A system can pass every validation check and still drift. Behavioral drift is not caused by invalid steps. It is caused by valid steps interacting in ways we did not anticipate. Increasing validation does not eliminate this problem. It only shifts where it appears, often pushing it further downstream, where it becomes harder to detect and correct.
Coordination becomes the system
If correctness does not compose automatically, then what determines system behavior? Increasingly, the answer is coordination. In traditional distributed systems, coordination refers to managing shared state, ensuring consistency, ordering operations, and handling concurrency. In autonomous systems, coordination extends to decisions.
The system must coordinate:
- Which information is used
- How that information is interpreted
- What actions are taken
- How those actions influence future decisions
This coordination is not centralized. It is distributed across models, planners, tools, and feedback loops. In agentic AI architectures, this coordination spans model inference, retrieval pipelines, and external system interactions. The system’s behavior is not defined by any single component. It emerges from the interaction between them.
In this sense, the system is no longer just the sum of its parts. The system is the coordination itself. Failures arise not from broken components, but from the dynamics of interaction timing, sequencing, feedback, and context. This also explains why small inconsistencies can propagate and amplify. A slight mismatch in one part of the system can cascade through subsequent decisions, shaping the trajectory in ways that are difficult to anticipate or reverse.
Control planes introduce structure, not assurance
One response to this complexity is to introduce more structure. Control planes, policy engines, and governance layers provide mechanisms to enforce constraints at key decision points. They can validate inputs, restrict actions, and ensure that certain conditions are met before execution proceeds. This is an important step. Without some form of structure, it becomes difficult to reason about system behavior at all. But structure alone is not sufficient.
Most control mechanisms operate at entry points. They evaluate decisions at the moment they are made. They determine whether a particular action should be allowed, whether a policy is satisfied, and whether a request can proceed. The problem is that many of the failures in autonomous systems do not originate at these entry points. They emerge during execution, as sequences of individually valid decisions interact in unexpected ways. A control plane can ensure that each step is permissible. It cannot guarantee that the sequence of steps will produce the intended outcome. This distinction is subtle but important: control provides structure, but not assurance.
From events to trajectories
Traditional monitoring focuses on events. A request is processed. A response is returned. An error occurs. Each event is evaluated independently. In systems exhibiting behavioral drift, behavior is better understood as a trajectory. A trajectory is a sequence of states connected by decisions. It captures how the system evolves over time. Two trajectories can consist of individually valid steps and still produce very different outcomes. One remains aligned. The other drifts. This represents a shift from failure as an event to failure as a trajectory, a distinction that traditional system models are not designed to capture.
Correctness is no longer about individual events. It is about the shape of the trajectory. This shift has implications not just for how we monitor systems, but for how we design them in the first place.
Detecting drift and responding in motion
If failure manifests as drift, then detecting it requires a different set of signals. Instead of looking for errors, we need to look for patterns:
- Changes in how similar situations are handled
- Increasing variability in decision sequences
- Divergence between expected and observed outcomes
- Instability in response patterns
These signals are not binary. They do not indicate that something is broken. They indicate that something is changing. The challenge is that change is not always failure. Systems are expected to adapt. Models evolve. Data shifts. The question is not whether the system is changing. It is whether the change remains aligned with intent. This requires a different kind of visibility, one that focuses on behavior over time rather than isolated events. Once drift is identified, the system needs a way to respond. Traditional responses, restart, rollback, stop, assume failure is discrete and localized. Behavioral drift is neither.
What is needed is the ability to influence behavior while the system continues to operate. This might involve constraining action space, adjusting decision selection, introducing targeted validation, or steering the system toward more stable trajectories. These are not binary interventions. They are continuous adjustments.
Control as a continuous process
This perspective aligns with how control is handled in other domains. In control systems engineering, behavior is managed through feedback loops. The system is continuously monitored, and adjustments are made to keep it within desired bounds. Control is no longer just a gate. It becomes a continuous process that shapes behavior over time.
This leads to a different definition of reliability. A system can be available, responsive, and internally consistent—and still fail if its behavior drifts away from its intended purpose. Reliability becomes a question of alignment over time: whether the system remains within acceptable bounds and continues to behave in ways consistent with its goals.
What this means for system design
If behavior is trajectory-based, then system design must reflect that. We need to monitor patterns, understand interactions, treat behavior as dynamic, and provide mechanisms to influence trajectories. We are very good at detecting failure as breakage. We are much less equipped to detect failure as drift. Behavioral drift accumulates gradually, often becoming visible only after significant misalignment has already occurred.
As systems become more autonomous, this gap will become more visible. The hardest problems will not be systems that fail loudly, but systems that continue working while gradually moving in the wrong direction. The question is no longer just how to build systems that work. It is how to build systems that continue to work for the reasons we intended.
AI & ML, Commentary
Radar
