Core Thesis: Purposeful, adaptive action requires more than raw data or environmental context—it requires a situation. A situation is the minimal, actor-relative composite structure upon which intelligent systems ground judgment, projection, and control. Modern observability has over-engineered the telemetry/evidence layer while completely outsourcing situational recognition to exhausted human operators.
Every intelligent system lives inside a changing world. Actors move. Resources are consumed. Relationships form and dissolve. Operations unfold and leave effects behind. Change arrives unannounced and unlabeled: another actor approaches, a resource disappears, a constraint tightens.
The observer encounters fragments of this unfolding world and attends to them selectively.
Purpose directs attention toward what may matter. Some fragments arrive as immediate signals. Others persist as traces of what has already happened. Together they form the evidence from which an intelligent system constructs its understanding of where it stands.
Intelligence begins with situations.
Evidence reaches the actor as signals; judgment and composition make that evidence actionable.
Situation as an Engineering Concept
Engineers are comfortable talking about state. They are comfortable talking about context. Situation evokes military command, emergency response, and social judgment. The concept has a rigorous literature behind it, most notably Endsley’s model of situation awareness developed for aviation, but it has largely remained confined to the domains that produced it, treated strictly as a human capacity.
That separation has an enormous cost, visible in the systems we build today.
Consider a threshold alert. It carries a single fact: latency exceeded two hundred milliseconds. That fact is equally consistent with a failing downstream dependency, a rolling deployment, a scheduled batch job running every Tuesday, or an inconsequential blip. The system records a state of affairs, and interpretation falls entirely to whoever reads the page at three in the morning. Measurement is automated; recognition remains manual. Multiply that by a few thousand alerts and the result is the familiar pathology of modern operations: abundant evidence, exhausted responders, and no machine-legible answer to the question that determines the response.
Situations are fundamental to goal-directed behavior. Every adaptive system continuously answers some version of the same question: What situation am I in? The answer is what grounds the next action.
Situation holds this central position because purposeful action requires a coherent composition of what is currently believed to be true. A situation composes evidence, judgment, and purpose into the smallest whole against which projection and choice become possible. A status can carry meaning on its own, much as a word can.
But judgment about what is happening requires composition: in language, assertion operates over something sentence-like; in situated intelligence, purposeful judgment operates over a situation.
The architecture developed here can be stated compactly:
Signal (Stimulus) → Sign → Status → Situation → Projection → Affordance → Action
Each stage structures uncertainty around what matters and moves the actor closer to purposeful intervention.
This architectural claim asserts that the same question arises at every capability level.
A thermostat approximates the same architecture in its thinnest possible form: one variable, two conditions, and evidence that settles the answer completely. What varies across systems is the difficulty of reaching the answer. A bacterium, an immune system, a distributed controller, and a human operator each encounter evidence that fits multiple scenarios simultaneously; each must choose an interpretation while staying open to error.
The architecture is universal.
The World Contains States of Affairs
The world holds states of affairs. A state of affairs is the arranged condition of the environment: the relationships among actors, available resources, existing commitments, physical constraints, and the consequences of prior events.
Engineering usually treats state as internal variables—memory, configuration, counters, or fields—representing only a narrow slice of the world’s full condition at a given time.
- A database replica is unavailable.
- A bounded queue is full.
- Network latency has doubled.
- Inventory has fallen below demand.
Each is an objective state; meaning requires an actor.
Signals, Signs, and Status
Observers notice evidence. Available information alone conveys states of affairs to them.
Every operation perturbs the environment. Some perturbations exist only while the operation runs; others outlast it.
A database transaction emits lock contention, CPU consumption, network traffic, and cache churn while it executes. When it completes, it leaves committed rows, updated indexes, timestamps, and modified balances behind.
Signals relate to states of affairs as footprints relate to the animal that left them.
A footprint’s value lies entirely in what it reveals about the entity that made it. Telemetry holds the exact same standing: a trace from which a condition is inferred. Control theory formalized this arrangement decades ago as state estimation—recovering a hidden condition from noisy observations—and the framing transfers naturally to software.
This is also why the present is legible at all. Operations pass through the world transiently and leave effects that persist. Any current state of affairs is the accumulated residue of countless completed operations.
The environment is a history written by action, and these memory traces are what make that history readable.
Operations occur in the world; operational representation is merely one way of observing them. Intelligence usually works from the effects, signs, and changing situations that operations produce. Detailed mechanism becomes important only when expectation fails or direct intervention requires it.
Evidence acquires qualitative significance through judgment, and that judgment occurs prior to situation recognition.
A rising queue is judged growing; continued accumulation makes it saturating, and eventually saturated. A dependency producing increasing failures is judged degrading.
These are signs—compressed qualitative judgments drawn from patterns in evidence.
Attributed to a subject at a point in time, a sign becomes a status: the queue is saturated, the dependency is degraded, the region is recovering. A sign specifies what quality has been judged; a status specifies what that quality is true of right now. A status can persist, strengthen, weaken, or mutate as further signs arrive.
Situations are composed of these judged conditions and the relationships among them.
- A bounded queue is full is evidence.
- Saturated is a sign.
- The queue is saturated is a status.
- Checkout capacity is exhausted during launch is a situation.
In the simplest controllers, status and situation collapse almost completely. A thermostat’s relevant world is so constrained that below setpoint already determines the actionable situation; the sign can trigger the response directly. But as relevant conditions, actors, and possible explanations multiply, status alone ceases to be sufficient. Purposeful choice among competing actions requires a situation.
Situation Is Actor-Relative
A situation is an actor-relative composition of statuses, relationships, and conditions, interpreted against that actor’s objectives, capabilities, and expectations. It is the actor’s reading of the relevant state of affairs.
Because interpretation is actor-relative, one world produces many simultaneous situations.
A congested network registers as packet loss for a transport protocol, as an exploit window for an attacker, as billed bandwidth for an infrastructure provider, and as launch-day success for the product team.
The state of affairs is shared; the significance varies with the actor.
A situation is necessarily partial. An actor holds only a fraction of the information available about the world, and intelligence must operate within that constraint: maintaining a representation sufficient for deciding what matters and what can be done. Biological perception and planning exhibit a similar economy, operating through preferential access to task-relevant information. A situation is therefore a functional compression, reducing high-dimensional observations to a compact relational structure that preserves what is consequential for judgment and action.
Situation is relational, constituted where an objective state of affairs meets a purpose.
Remove the purpose, and only the arrangement remains. Barwise and Perry established a precedent in semantics for treating situations as first-class partial structures, irreducible to collections of isolated facts. Their situations are actual fragments of the world; the usage here adds an explicitly actor-relative dimension, where situation is the significance such a structure acquires for an agent endowed with specific purposes and capabilities.
This distinction is what separates situation from context, two terms frequently conflated.
A latency metric carries context: service name, deployment version, region, workload class, customer tier, time of day, upstream caller. Context supplies the environmental conditions within which an observation is interpreted.
Situation is the actor’s integrated reading of what those conditions signify for its goals, capabilities, and expectations.
Modern software infrastructure has become remarkably adept at understanding context. This includes tags, topologies, service meshes, deployment metadata, and distributed traces. These elements collectively enhance interpretation without actually performing it.
Context provides the framework for interpretation, while the situation grounds judgment and control.
Recognition Happens Under Uncertainty
Recognition can be mistaken, and error handling is an intrinsic part of the design.
Much recognition is computationally inexpensive because an actor carries a running expectation of the situation it occupies. Incoming signs continually confirm, weaken, or violate that expectation. As long as observation remains aligned with expectation, the situation requires little active reconstruction. Surprise is what forces recognition into the foreground: competing hypotheses become salient, attention widens, and the actor may descend into underlying mechanisms to explain what its existing model can no longer account for.
Expectation enables partiality: the actor assumes what can be safely ignored until signs suggest otherwise.
The rustling grass may be wind, or it may be a predator. Evidence alone is insufficient to determine the exact situation; multiple scenarios can often be consistent with the same signals. An actor therefore maintains several competing situation hypotheses, each with an associated confidence. Recognition occurs when a hypothesis gains enough confidence to justify a decision, considering the consequences of error. Since recognition involves commitment in uncertain circumstances, the fundamental engineering question is which failure modes a system should prioritize.
Those costs are profoundly asymmetric. Mistaking wind for a predator costs a few seconds of wasted flight; mistaking a predator for wind costs an organism its life. Natural systems are tuned accordingly, and well-designed controllers must be as well: an autoscaler that occasionally over-provisions incurs a minor financial penalty; one that under-provisions during a major product launch risks systemic collapse. Recognition is a decision with asymmetric costs. Any architecture that treats it as simple classification can be mathematically well-calibrated but still behave disastrously.
A deeper problem sits underneath: an actor can recognize only the situations it possesses the means to represent. That vocabulary of situation types must come from somewhere—evolved, learned, configured, or inherited from an operator’s runbook. A situation vocabulary holds more than identifiers; it encodes types and the relational constraints connecting them: which configurations of sign and status support a given type, which situations imply or precede others, and which transitions become plausible once a type is recognized. Situation theory makes this explicit in the canonical case of recognizing a smoke-type situation while attuned to the constraint linking it to a fire-type situation. Recognition is possible because the actor has encoded regularities across situations.
A system with rich telemetry and an impoverished situation vocabulary will misclassify with absolute confidence.
Projection Extends Recognition
Recognizing the present is only half of intelligence. Control and steering requires anticipating its future.
Projection selectively explores possible trajectories from the current situation, conditioned by the observer’s model, objectives, and immediate concerns. Because its ultimate aim is intervention, projection deals specifically in futures that are both plausible and still alterable by actions taken now. Projection is selective for the same reason perception is: intelligence directs computational effort exclusively toward futures that matter.
Projection operates in two distinct modes:
- Trajectory projection advances the situation forward under existing dynamics, asking where the system is heading without intervention. This establishes whether action is warranted.
- Counterfactual projection advances the situation forward under candidate actions, asking where a particular intervention leads. This makes alternative interventions comparable.
An autoscaler demonstrates both modes: trajectory projection and counterfactual projection. In trajectory projection, when the arrival rate surpasses the service rate, the queue continues to grow, leading to overload. Conversely, in counterfactual projection, adding four worker nodes increases the service capacity beyond the arrival rate, enabling the queue to drain. Shedding low-priority workloads immediately reduces the arrival rate and restores latency bounds.
Projection operates on both sides of affordance discovery: trajectory projection identifies the need for intervention, affordances enumerate the available options, and counterfactual projection evaluates each alternative.
This explains why early transformation recognition is so valuable.
An operation is a transformation in progress—one state of affairs mutating into another. Recognizing a shifting situation while that transformation is still underway preserves the window of intervention. The rustling grass matters precisely because it projects forward to the predator’s arrival. Projection gives control its temporal horizon, and that horizon is bounded by the validity of the underlying model. Structural shifts in the environment can outrun standing projections, requiring the system to recognize when the model itself demands revision.
Affordances Are Possible Transitions
Every environment permits numerous actions, but their significance varies by actor.
Drawing on Gibson’s ecological framing, affordances can be represented computationally as the state transitions available to a specific actor, endowed with specific capabilities, from a specific situation.
A rock affords shelter to a rodent, elevated vantage to a hawk, and building material to a human. The rock remains constant; the affordances emerge from the pairing of capability and context.
Affordances arise from the intersection of situation and capability.
Objectives determine which available transitions are relevant. Goals define desired states of affairs, and operations are selected from relevant affordances because their projected outcomes move the world toward those desired states.
Affordances sit squarely at the frontier between recognition and control:
Given the situation I believe I am in, what transitions are available to me?
The qualifier holds significant weight: the actor responds to the situation they perceive themselves in. Controllers operate based on their current situational awareness, which is why intelligence requires more than just measurement.
Because affordances depend on capability, the same situation yields different action spaces for different actors. This is where the boundary between automated execution and human judgment naturally falls. An autoscaler and the supervising engineer can share a recognized situation while holding entirely different sets of affordances:
- The autoscaler can add compute capacity or adjust concurrency limits.
- The engineer can roll back a release, invoke a circuit breaker, fail over to another cloud provider, shed an entire customer class, or declare an incident.
The recognized situation serves both consumers simultaneously: the automated controller selects from the subset of transitions it is authorized to execute, while the human operator selects from a wider, more strategic set.
The situation is shared; the affordances diverge.
The operational line, a crucial design decision, determines when an autonomous system should act and when human intervention is necessary. Autonomy should be granted where recognition confidence is high, consequences are manageable, and actions are reversible. However, as uncertainty increases or consequences become irreversible, escalation to human intervention is essential. Architectures that base this boundary solely on technical capability will inevitably execute decisive actions in situations they lack the authority to resolve without supervision.
Acting Changes the Situation
Every intervention immediately starts to erode the recognized situation that justified it. This dynamic is easily obscured when feedback is depicted as a neat loop (Recognize → Project → Act → Repeat), because diagrams imply a linear sequence of events. In reality, an action begins to change the world even before the recognition that prompted it is fully utilized. For instance, an autoscaler adds additional nodes; the situation it was responding to promptly ceases to exist, replaced by a new state where compute capacity is being initialized, financial costs are increasing, and the initial queue saturation may or may not have cleared. Recognition must then begin anew in an environment that has been altered by the controller’s own output.
Other autonomous actors compound this volatility. Every external actor operates its own loop against the same shared environment on its own schedule. Their actions arrive as unannounced revisions to a recognized situation already in flight. This is where reflexivity enters. In financial markets, during incident mitigation, or within multi-agent distributed systems, acting on a situation alters the situation for every other actor observing the same world.
The practical consequence is that every recognized situation has a finite half-life.
Control always acts on a representation that has already begun to expire. Decay rates vary by orders of magnitude: an infrastructure topology may remain valid for hours, a market liquidity situation for microseconds, an incident hypothesis during a cascading failure for seconds. Systems remain stable only as long as their control loops cycle faster than situations decay. When decay outpaces renewal, controllers act on states of affairs that no longer exist.
Many outages recorded as “bad decisions” were entirely sound decisions applied to a world that had already moved on.
Recognition is delayed, while projection advances the recognized situation toward the moment of intervention.
Configuration management systems effectively demonstrate this dynamic with remarkable simplicity. They meticulously inspect the running system to estimate its current state, compare it against a predefined target, compute the difference, and execute operations specifically designed to reconcile the discrepancy. This process is automated, explicit, and completes in a single loop. As a result, the system’s state is entirely flattened into a static representation. Since the desired conditions are statically declared, the difference acts as a proxy for recognition. This is precisely why such tools become vulnerable when the environment deviates from their underlying assumptions. A drifted configuration file and a compromised host can produce identical differences. Consequently, a system that lacks a situation layer responds to both by blindly rewriting the file.
What Observability Is Missing
Everything described here has been understood, across various disciplines, for decades. The gap lies in what our industry actually constructs. We have engineered the evidence layer with extraordinary sophistication (metrics, structured logs, distributed traces, continuous profiling, topology graphs, runtime metadata) and then stopped.
Almost everything above the raw evidence layer is outsourced to human operators.
A controller responds to a recognized situation; evidence serves solely to inform and refine that recognition. However, the current practice of observability provides raw evidence and stops short of the boundary. Its architecture encourages the collection of increasingly large volumes of data while entirely delegating semantic interpretation to the engineer’s cognitive abilities. The synthesis of evidence into signs, signs into statuses, and statuses into situations occurs within someone’s mind at the time when they have the least capacity to do so.
Intelligent behavior operates through purposeful partiality:
- isolate what is consequential,
- synthesize a sufficient situational model,
- project trajectories,
- and act.
Reaching this standard necessitates the one element that current architectures lack: an explicit vocabulary of situation types, which should be maintained directly within the system rather than confined to runbooks and tribal memory. The rest of the process follows from treating the situation as a first-class software primitive—instantiated, updated, projected, and shared across automated controllers and human operators.
Observability, properly conceived, begins not with the collection of telemetry, but with the recognition of a situation.
The Architecture of Intelligent Behavior
The world contains states of affairs. Purpose directs attention, and actors receive partial evidence of those states as signals. Judgment compresses signals into signs; attribution turns signs into statuses. Statuses compose into competing situation hypotheses. Recognition commits to one with sufficient confidence to ground a decision. Trajectory projection explores where that situation is heading. Affordances expose the transitions available to this specific actor. Counterfactual projection evaluates candidate interventions. Control executes the chosen transition. Effects accumulate, the outcome is measured against expectation, the model is revised, and the cycle renews.
Purpose governs this architecture at every phase: it dictates which evidence commands attention, which situation types become salient, which futures are projected, and which affordances matter.
Situations compose hierarchically:
- A dependency failing under retry amplification is a situation.
- Retry storms overwhelming upstream capacity is another situation.
- A service approaching total thread starvation is a third.
Together, they form an incident situation—each remaining fully recognizable at its own level of abstraction.
Higher-order situations are composed from lower-order ones, preserving the relational structure that maintains the overall actionability. This composability is what makes situations the fundamental unit of intelligence: they are the minimal coherent structure upon which purposeful action can be based. Additionally, it enables the architecture to scale, preserving identical semantics from localized control loops to organizational cognition.
This architectural pattern is found in various fields, including biology, cybernetics, organizations, markets, and autonomous software, as it forms the core structure of purposeful behavior.
Discussions of intelligence frequently become trapped in data because data is tangible, measurable, and easily monetized. Situations are demanding. They exist only at the intersection of an actor, its objectives, its capabilities, and the dynamic world it inhabits. Every intelligent action is grounded in a situation.