For many, traces seem to be a fundamental building block of telemetry. Computers operate through a series of calls, with each procedure triggering the next, resulting in a cascading chain of invocations. When telemetry is organized around traces, it appears to be an intrinsic part of the machine’s architecture. Tracing essentially records the machine’s calls. While this intuitive understanding is helpful, it falls far short of providing a comprehensive grasp of systems.
Hold that intuition next to a brain. A brain operates on firings. A neuron spikes, pushing others past threshold, creating a cascade of firing. This firing is real, the substrate of thought, as fundamental to cognition as computation.
For a time, neuroscience aimed to record firings, correlate them with thought, and find the mind in the machine. A single-unit recording indicates a cell spike, but it doesn’t reveal the thought behind it. Stacking and time-aligning recordings shows activity filling a raster, but the thought doesn’t appear because it wasn’t at that level. The explanatory level existed within the circuitry, representing activity and its meaning to the system. Firing is the substrate of mind, not the mind itself. Confusing the two cost the field years before it reached the level where the question arose.
Neuroscience has shown us that no matter how detailed we get in understanding the underlying firing structure, the real understanding happens at a much higher level. The ghost in the machine.
Observability is a property that measures how much a system’s internal state can be inferred from its outputs. Control theorists defined it as the ability to reconstruct a system’s full internal state from its outputs over a finite interval. This requires a state space to reconstruct into. Without a model, it’s impossible to recover a system’s state. Coordinates are meaningless without a manifold. Record without a model and you’ll have spikes with no place to put them.
Situational awareness involves perceiving elements, understanding their meaning, and projecting their future. Telemetry attempts but fails to support the first step. The traces are recorded, timestamped, and stored without signs or significance. Comprehension seeks a model of their combined meaning, while projection requires a stable model for future prediction. More data raises the floor but leaves the ceiling fixed, causing overload with no structure to make sense of the signals. Under this overload lie three crucial questions: what does the system do, what state is it in, and how is it steered? Each presupposes a model of the system as a thing with behavior, state, and inputs that move it. The collector defers these questions, and the deferral is comfortable because the model, when asked directly, is missing.
Where does the model originate? We must transcend yesteryear approaches and reach for the level where activity gains meaning, which is the same the shift neuroscience made when it abandoned reading minds from single cells.
Medicine began with diagnosis, which involves interpreting signs. A fever, for instance, is a value against a pathology model, conveying something to the reader. The symptom itself isn’t the disease; it’s what the disease presents to the reader. The physician’s expertise lies in the interpretant, transforming a mark into a referent. Without this skill, accurate measurements lose their value, much like a dashboard showing an endless array of tree nodes within a trace or an unending list of log records with hundreds of fields and tags, falsely claiming to be semantically meaningful.
Biology used signs when mechanism failed. Living systems coordinate by interpreting signs, not forces. A cell responds to a molecule based on its meaning to its machinery, not its mass. The grammar of mechanism couldn’t handle complexity, so biology used the grammar of signs. A system whose parts coordinate by interpreting each other needs a sign-theoretic account. A distributed software system is exactly that: services acting on messages based on meaning.
The digital system is mechanically simpler than the brain. Its substrate is legible, deterministic, and inspectable in ways that neural tissue never allows. We can capture every firing perfectly. However, in one respect, it is more challenging because the brain is pre-programmed with its meanings, while the digital system was constructed by us. We never provided a vocabulary for its signs. We built the firing process but skipped the language. The system’s purpose, the states it seeks, and what it considers well-formed were never given a vocabulary and hence never expressly shared.
Then agents arrived, further exasperating the error failing of tracing. An agent sits where neither available vocabulary can reach it. The function vocabulary calls it input and output — a call, tokens returned, latency logged. Electrodes again. An agent responds to a stimulus by interpreting it, while built as a function, carrying a human’s intent into action. It lives between a function and an extension of human agency, interpreting inward and pointing outward to a principal.
Reading agents through their human language is ineffective because it’s verbose, ambiguous, and generated rather than measured. You can’t understand many agents by reading the chats of each, just as you can’t understand a market by reading every trader’s notebook. Understanding at scale requires a typed sign system with a controlled vocabulary where each act registers as a sign of a known kind, combinable by a grammar, and legible without the prose.
An agentic system must rely heavily on a sign-based system for sense- and meaning-making. This requires a controlled vocabulary of signs, a syntax for their combination into states, and an interpretant process .When these components are present, the system’s behavior becomes legible and a shared situational awareness becomes possible for all.
Neuroscience found the mind only after it stopped expecting the spike to contain it and climbed to the level where firing becomes representation. The work in front of observability is the same climb. We have the fired emissions.
What remains is the language they were always supposed to be speaking.