Core Thesis: AI agent architectures are semantically inefficient because they constantly serialize internal states into natural language prose only to pay probabilistic models to re-interpret them. To achieve scalable, intelligent behavior, AI must transition from purely linguistic systems to semiotic agents—architectures that convert expensive linguistic reasoning into persistent operational distinctions (“semiotic compression”), reserving rich language processing only for novel, ambiguous, or high-consequence situations that cross a threshold of significance.
Internal Language Creates Semantic Inefficiency: Using natural language prose as the primary protocol between agent components forces systems to repeatedly dissolve concrete operational facts into text, only to spend probabilistic compute reconstructing them downstream.
Semiosis Runs Deeper Than Narration: Meaning occurs whenever a situation, relationship, or even silence changes what an agent can do next based on expectations and context. True agency relies on processing these sign-relationships, not just generating sentences.
Language Belongs at the System Boundary: Natural language excels at handling ambiguity, negotiating novel rules, and communicating with humans. Internal system boundaries, however, should retain established operational distinctions directly without re-translating them into prose.
Learning Operates via Semiotic Compression: Expertise shortens the path from perception to action. Learning compiles laborious multi-step reasoning down into direct, low-cost operational competence, transforming deliberate inference into immediate recognition and action.
Semiotic Thresholds Force Decompression: Significance is situated. When a situation encounters novelty, unexpected resource conflicts, or heightened risk, it crosses a threshold that signals the agent’s compressed habits are insufficient, triggering a shift back to explicit, high-cost reasoning.
Attention Is an Allocation Problem: Compute, latency, and context windows are finite constraints. Capable agents must treat attention as a dynamic resource, allocating expensive inference, simulations, and human consultations only to situations that warrant them.
Agency Requires Dual-Channel Governance: Systems must run an activity flow to accomplish tasks alongside a parallel regulatory flow to continually evaluate whether the system remains authorized, confident, and competent to proceed in the current context.
Human Control Belongs Above the Threshold: Scaling autonomous systems requires positioning human operators above the threshold of significance. Humans intervene when a situation exceeds an agent’s authority or competence, optimizing operator attention while maintaining safety.
Authority Is Relational, Not Syntactic: Security risks like prompt injection arise from treating authority as raw text. Legitimate instructions must be evaluated relationally across the source, situation, existing commitments, and logical permissions.
Observability Requires Operational Signs, Not Text Traces: Reading exhaustive internal logs becomes unviable as agents scale. Effective monitoring requires high-level signs that expose the agent’s current state: recognized commitments, working assumptions, uncertainty levels, and active risk thresholds.
Making meaning in a world of consequence
Current AI agents are strangely verbose.
Give an agent a task, and much of what happens beneath the interface is language addressed to language. A model interprets a prompt, produces a plan, calls a tool, receives another representation, updates a textual or structured context, consults a memory, and invokes another model. Communication between agents frequently consists of one language model generating tokens for another language model to interpret. This works well.
Language is the most general interface humans have developed. Large language models inherit an extraordinary amount of conceptual structure from it, and natural language gives people, models, software and tools a convenient layer of interoperability. It is worth asking, then, whether an such useful interface has come to stand in for the architecture of intelligence itself.
Human competence runs largely beneath narration. We learn distinctions. We recognize situations. We acquire habits. We anticipate. We respond to gestures, absences, changes, relationships and expectations. Much of what matters to us operates without ever becoming a sentence. Language is one of our capacities. The wider capacity is semiosis.
We are semiotic agents. Increasingly capable computational agents need to become semiotic agents too.
Semiosis as process
The word semiotic here names a process. Symbols, icons, agent languages, event taxonomies and ontologies can all serve that process; semiosis is the activity in which they come to matter. A sign matters when something is taken to stand for something else, in a context, by an interpreting process, with consequences for what can happen next.
Consider something very simple: a message fails to arrive. The channel is silent. There is no token to classify. Yet the silence can become a sign. If nothing was expected, the silence is simply silence. If a message was expected eventually, the silence may indicate delay. If a commitment required the message before another action could proceed, the same silence may indicate that the commitment is at risk. The physical situation is identical in each case. Its significance is relational: it depends on expectations, context, history, purpose, time, and commitments. Once interpreted, the silence changes what the agent does. It waits, investigates, inhibits another action, reallocates a resource, asks someone, or revises an expectation. That whole process is semiosis.
Definition: A semiotic agent is a participant in an environment that can develop, use, and revise relationships between signs, situations, expectations, and possible actions.
The emphasis falls on the verbs. Meaning is something the system continually does.
Where the distinction lives
This framing exposes a weakness in many current agent architectures.
Enterprise agent systems move enormous quantities of representation between components. Prompts, retrieved documents, conversation histories, summaries, plans, tool responses, embeddings, and “memories” are assembled into context so that another model can infer what is happening. Representation is necessary everywhere.
An opcode is a symbol; so are the contents of a network packet, an event, a database value, a protocol message.
Question: Where does the distinction that matters actually live?
Suppose one agent sends another: The customer appears unable to complete payment. What is that? An observation? A hypothesis? A diagnosis? A warning? A state transition? A request for intervention? An instruction to halt fulfillment?
The sentence can serve any of those functions, and its operational significance has to be reconstructed from context. That flexibility is one of language’s extraordinary properties and a large part of why language models are so useful. Placed at every internal boundary of a computational system, the same flexibility becomes a recurring cost.
When a system already knows that a payment commitment has expired, that a dependent operation must now be inhibited, and that an alternative settlement process is available, serializing those distinctions into prose and asking another probabilistic model to recover them can amount to semantic inefficiency. The architecture dissolves operational distinctions into representation and then pays to reconstruct them.
A semiotic architecture keeps representation and preserves distinctions once they have become meaningful.
Language at the boundary
Humans are linguistic beings, and language excels wherever meaning is uncertain, contested, novel or socially negotiated. A person can say: “I know this normally counts as completed, but in this case the learner was assisted throughout the assessment.” That utterance introduces a distinction the computational system has yet to hold.
Language lets us explain exceptions, negotiate intentions, challenge interpretations, introduce concepts, communicate experiences and repair misunderstandings. Once a distinction is operationally established, the system can hold it directly and spare itself the work of rediscovering it from prose on each occasion.
The architecture therefore runs in two directions:
language → interpretation → operational meaning
operational meaning → explanation → language
Language remains the principal human–machine interface, and internal acts of agency run through established operational distinctions. The result is linguistic intelligence used where language has comparative advantage.
Learning compresses interpretive work
The most important property of semiotic agency is that learning changes the cost of interpretation.
Compression here refers to the interpretive path between encountering a situation and becoming appropriately oriented toward it. Learning shortens that path.
Three forms of compression merit distinction:
- Information compression asks how economically something can be represented.
- Computational compression asks how economically a result can be calculated.
- Semiotic compression asks how much interpretive work experience can make unnecessary while preserving sensitivity to consequential differences.
The third is the concern here.
Consider learning to drive. For the novice, every situation is expensive. Mirrors, steering, road position, signs, other vehicles and pedestrians each demand attention, and many relationships have to be consciously reconstructed.
The experienced driver notices more. The slight drift of another vehicle within its lane, the posture of a pedestrian about to step into the road, the geometry of traffic forming ahead of an obstruction: these register without laborious reasoning. Experience has established relationships between signs, situations and consequences, and what once required explicit interpretation now participates directly in competent activity.
Learning has compressed the interpretive work while preserving, and often enriching, the discrimination.
Effective compression preserves action-relevant distinctions, enhances sensitivity to them, and minimizes effort.
Chess shows the same pattern. Expertise keeps the rules and the board fully in view and adds significance to familiar configurations. What a novice discovers through deliberate search, an expert can recognize as a meaningful pattern: pressure, vulnerability, structure, opportunity. The interpretive path has shortened.
Semiotic agency is implementation-neutral.
A symbolic system can encode a shortcut as an explicit rule; a neural system can acquire representations that map situations to responses ever more directly. Either can carry semiotic compression, which is a functional property.
Question: Has experience changed which distinctions carry significance, so that competent orientation requires less reconstruction while consequential differences keep the power to change that orientation?
An inexperienced agent might require:
retrieve → describe → compare → infer → explain → decide
With experience, a recurring situation increasingly supports:
recognize → orient → act
The intermediate reasoning is still accessible. It has grown unnecessary for this encounter because the agent has gained competence. Learning is the establishment of what an experience means for future action.
Question: What no longer requires reconstruction due to what the agent has learned?
The semiotic threshold
A semiotic threshold is a threshold of significance. It marks the point at which something about a developing situation means the agent’s existing way of handling it may be inadequate.
Question: Is my current interpretation sufficient for what is becoming consequential?
An entirely normal observation can cross that threshold. Suppose two independent plans remain individually valid. Every service is functioning, and every metric sits within range. The plans now require the same exclusive resource at the same time. A relationship has changed significance, and the threshold has been crossed.
An unusual event, equally, can sit well below the threshold when it bears on nothing the system is trying to accomplish.
Novelty, anomaly and uncertainty can each contribute to significance. Significance itself exists relative to a situation, a purpose and a possible consequence, and semiosis is what establishes that relationship.
A learned interpretation is economical because only some distinctions participate in each act. Significance, however, is situated. A distinction irrelevant across a thousand encounters may become decisive in the thousand-and-first because the purpose, context, relationship, authority, timing, or consequences have changed. Semiotic compression therefore has to remain defeasible. Distinctions set aside from the economical path stay recoverable, and experience keeps the power to challenge the very distinctions through which the agent currently interprets experience.
This gives the semiotic threshold its deepest role. It detects that: The current economy of interpretation may be inadequate to the situation. At that point, distinctions held outside the economical path are recovered. More evidence is examined. A broader context is considered. Another agent participates. Expensive reasoning resumes.
Learning operates in both directions. It compiles interpretation downward into competence, and changing significance decompresses competence back into interpretation: interpretation → learning → competence → economy, and, when adequacy is challenged: significance → threshold → renewed interpretation.
The objective is the minimum interpretive work consistent with preserving important distinctions.
Attention follows significance
Once significance crosses a threshold, the agent’s resourcing changes. More evidence might be gathered, a longer history examined, another interpretation considered, a simulation run. Another agent might be involved, a larger model invoked, an action made reversible, authority reduced, a person consulted. The threshold therefore answers a second question: What kind and amount of interpretive resource does this situation now warrant?
Attention becomes an allocation problem.
Humans appear to operate under this constraint because cognition is expensive. Competence pushes enormous amounts of activity below deliberate attention, keeping scarce cognitive resources available for situations that warrant them. Computational agents face constraints of their own. Inference costs money. Latency matters. Context is finite, and human attention far more so. Simulation, communication, and coordination all carry cost. Additional reasoning can also degrade judgement. A capable agent learns where additional interpretation is worth its cost.
Nested timescales of interpretation
The familiar distinction between fast and slow cognition points toward a richer structure.
Distributed systems offer a useful analogy in data planes and control planes: one part of a system performs activity according to an established organization, and another regulates, configures, or changes that organization. Biological systems contain many analogous separations operating across different timescales.
The governing principle is that performing activity and regulating how activity is performed are distinct functions.
For agents, an activity flow:
encounter → interpret → coordinate → act
runs alongside a regulatory flow:
significance → uncertainty → authority → attention → escalation
The two interact continuously, and each can be distributed across many components. A local process might recognize a familiar condition in milliseconds. A wider process might reconsider relationships across a larger situation. Another might revisit assumptions over hours. A human organization might reconsider the purpose itself over months.
Fast and slow at one level form a nested hierarchy of interpretive timescales across the system.
Control begins with interpretation
Classical control begins with deviation:
desired state − observed state → error and acts to reduce the error.
Intelligent systems face a prior question:
What counts as a deviation here?
A temperature of 90°C gains meaning from its context. Its importance depends on the material being heated, the current phase of operation, the expected next step, the duration of the condition, and what relies on it.
Before control can respond to a difference, some process must establish the significance of the difference.
The loop becomes: difference → interpretation → significance → constraint → response
Semiosis sits naturally upstream of adaptive control.
Because interpretation can address different temporal relationships, the same architecture supports more than reaction. A sign can indicate what has happened, characterize what is happening, and suggest what is becoming likely. Reaction, recognition, prediction, and projection are temporal orientations of a single interpretive process.
The regulatory channel of agency
An autonomous agent continually faces two questions: what should I do? and am I still entitled and competent to do it?
An agent can operate entirely within its technical permissions and still misunderstand the situation. Database access, authorization to update records, and a valid objective are each preconditions. Warrant for a specific action depends additionally on whether the agent’s present interpretation of the situation holds.
This suggests a control channel running alongside agency. The activity channel accomplishes work. The regulatory channel maintains the conditions under which that work remains legitimate. A failure has two interpretations.
For the activity process: failure → diagnose → recover
For the regulatory process: unexpected failure → confidence reduced → consequence increasing → delegated competence approaching its boundary
The activity process asks how to continue. The regulatory process asks whether continuing in the same mode remains justified. Control becomes partly constitutive of agency, woven into the agent’s own interpretation of its situation.
This arrangement defines a clear point for human involvement: above the threshold. A person steps in when a situation grows beyond the agent’s assigned competence, authority, confidence, consequence, or ability to recover. Such positioning maintains both the scalability of autonomous systems and the speed of human control.
Thresholds can be situated and learned: novices escalate often, but experience builds competence. Competence compresses interpretation, freeing reasoning and attention from routine cases.
Learning must also enhance an agent’s ability to recognize when its competence no longer applies. A capable agent knows what to do and when its knowledge reaches its limit.
Human control scales as attention shifts to situations where interpretation matters.
Subordination requires interpretation
The idea that artificial intelligence should stay aligned with human goals seems straightforward until an agent operates beyond a short span—from seconds down to milliseconds or less.
A human gives it a purpose. Then the world changes. Resources disappear. Other participants act. Plans conflict. Evidence contradicts assumptions. Consequences emerge that were invisible when the instruction was given.
Faithfulness to the original intention requires the agent to determine, continually, whether its evolving actions still instantiate the delegated purpose. That is an interpretive problem.
Subordination itself requires ongoing semiosis:
purpose ↔ situation ↔ interpretation ↔ action ↔ consequence
Question: How does an autonomous system continuously interpret its authority in its current context?
A policy constrains the answer. Interpreting the question remains the agent’s continuing work.
Consider the instruction: Delete the records. Encountered on a webpage, those words have one significance. In an email, another. Issued by an authorized operator through a governed control interface, another still.
The representation is identical. What changes is the relationship between the sign, its source, the situation, existing commitments, and the structure of authority. Prompt injection is therefore a semiotic problem at root.
An agent must distinguish four things: information about an instruction; an instruction; an instruction issued with legitimate authority; and an instruction it is permitted to execute in the present situation.
Each of these distinctions is relational, held between the text, its source, and the situation. Authority is semiotic.
This analogy gives the data plane and control plane real weight. Information and authority may move through the same physical channel while occupying distinct logical channels of meaning.
Observing through signs
This changes what it means to observe an agent. Current practice records everything: prompts, responses, tool calls, reasoning traces, memory retrievals, actions. Those records serve diagnostics and forensic reconstruction well.
Understanding agency calls for something further.
The more sophisticated the agent, the less plausible it becomes that an operator will understand it by reading everything it generated. What needs exposing are the distinctions governing its current participation:
- What situation does it believe it is in?
- What commitments does it recognize?
- What does it currently expect?
- What authority does it believe it possesses?
- What assumptions support its current interpretation?
- Where is uncertainty accumulating?
- What is becoming consequential?
- Which threshold is being approached?
- What would cause it to reconsider?
These are signs of the agent’s operational orientation.
A trace records what happened inside a computational process. A sign tells another participant what that process currently signifies for the shared activity. Observing agents calls for signs.
A semiotic environment
The deeper architectural implication is that semiosis belongs to the environment agents share.
Meaning is relational. An agent encounters an environment. Other agents act. Commitments are created. Expectations develop. Actions produce consequences. Interpretations become shared, disputed, corrected, and stabilized.
The object of design becomes the semiotic environment in which agents participate.
Such an environment lets occurrences become signs; lets signs acquire situated significance; keeps interpretations provisional; forms and tests expectations; lets commitments constrain future possibilities; lets consequences confirm or challenge interpretations; allows novel distinctions to emerge and useful ones to stabilize through experience; and propagates consequential uncertainty toward greater interpretive resources and authority.
Such a substrate can carry messages, hold context, retain memory, and include ontologies. Its defining role is to serve as the medium through which experience becomes consequential. Agents participating in it could contain language models. Some could be predominantly symbolic. Others could use learned latent representations. Some could be humans. What makes them participants in the same system is their ability to establish and negotiate enough shared significance to coordinate activity, whatever their internal representations.
Humans are semiotic agents too
Organizations are enormous semiotic systems. A signature changes the status of a document. A red traffic light reorganizes possible movement. A wedding ring signifies relationships that its material composition leaves entirely open. A flashing icon demands immediate attention from one operator and passes unnoticed by another. A manager saying “fine” can signify agreement, resignation, irritation, or the end of a discussion, depending on context.
Money is overwhelmingly semiotic. Law is overwhelmingly semiotic. Professional practice depends on learning which differences matter in situations that superficially resemble one another.
Expertise is partly the acquisition of a sophisticated field of significance. The experienced nurse, engineer, teacher, pilot or trader notices differently. Things invisible to the novice become signs. Things that overwhelm the novice become routine. Some situations immediately warrant attention; others can safely remain beneath it.
Semiotic compression is a matter of seeing more with less effort. The expert perceives more meaningful distinctions while spending less interpretive effort on familiar ones. Expertise changes both discrimination and the allocation of attention. That is precisely the capability computational agents need to acquire.
Machines will go further
Humans developed under severe perceptual, cognitive, social, and biological constraints.
Computational agents inhabit different environments. They can perceive relationships across timescales we struggle to hold in awareness, coordinate across enormous populations, and encounter computational phenomena for which ordinary language has poor concepts. They may develop distinctions whose usefulness is demonstrated through prediction and coordination long before we have convenient words for them.
The test lies in what the distinction allows. Does it forecast outcomes or transfer across settings? Does it aid coordination and endure shifting conditions? Might experience disprove it or keep meaningful differences intact? When effects reach the human sphere, can they become clear enough to guide action? Internal forms may adopt any shape suited to the machine. The demand rests at the edge: interfaces readable by people for signification of agency.
From prompts to participation
The current generation of agents has demonstrated something enormously important: language models give machines a remarkably general capacity to interpret representations. That capacity may prove to be a bootstrap.
The next architectural step lets interpretations produced through expensive linguistic reasoning become persistent operational distinctions. Experience refines those distinctions. Learning compresses their use. Semiosis continually tests their significance against changing situations. Thresholds determine when established competence suffices and when additional attention, computation, authority, or human participation is warranted.
Language remains available whenever meaning must be negotiated again.
The progression runs:
representation → interpretation → significance → disposition → action → consequence → learning
and, through learning:
experience → distinction → competence → interpretive economy
until something changes sufficiently that:
significance → threshold → renewed interpretation
This suggests a view of intelligence defined by participation: whether a system can engage effectively in a shifting environment, noticing what counts, discarding what can be safely ignored, foreseeing what might matter next, acting within its scope, learning from outcomes, and sensing when its current grasp has reached its boundary.
Humans have practiced this for a long time.
The potential of semiotic agents lies in machines that gradually assign meaning to their interactions with the world.