A trading bot keeps executing while the market regime changes. A warehouse robot keeps optimizing its route while a sensor slowly drifts. A customer-service agent keeps sounding fluent while the conversation loses coherence one turn at a time.
From the outside, the system still looks agentic. It acts. It responds. It may even keep producing acceptable short-term outcomes. The dashboard, naturally, waits until the mess is obvious. Dashboards are polite like that.
The paper A Mathematical Theory of Agency and Intelligence makes a sharper claim: the missing variable is not more reward, more benchmark performance, or more fluent prediction. The missing variable is whether the system can monitor the quality of its own coupling with the environment.1
That sounds abstract. It is not. The authors propose a metric called bi-predictability, written as $P$, which measures the fraction of total information in an interaction that is actually shared across the loop connecting observation, action, and outcome.
In plainer terms: how much of the system’s informational budget is doing real work in keeping it connected to the world?
That is the mechanism. The business implication follows later: if $P$ falls before reward falls, then operators get an earlier warning signal. If the fall can be decomposed into where predictability was lost, then the system gets not only an alarm, but a diagnosis. That is the difference between “something is wrong” and “the agent has lost grip in this specific way.” Subtle distinction. Expensive when ignored.
The paper’s core move is to measure the whole loop, not the score
Most AI reliability tooling watches fragments. Reward. Accuracy. Drift. Confidence. Human feedback. Semantic similarity. All useful. None necessarily tells us whether the agent’s internal state, its actions, and the resulting world state remain mutually constraining.
The paper defines predictive coherence for a passive system as:
Here, $S$ and $S’$ are successive states. The numerator measures shared information. The denominator measures the total information capacity of the two-state interaction.
For an agentic system, where an action channel enters the loop, the definition becomes:
where:
- $S$ is the state or observation available to the agent;
- $A$ is the selected action;
- $S’$ is the next state or outcome.
The important point is not merely that the agent has information. Everyone has information now. We have terabytes of the stuff lying around, mostly making people worse at meetings.
The important point is what fraction of the information deployed by the system is shared across the actual interaction. A system can process a large amount of information while only a small part of that information remains useful for predicting what its actions will do next.
That distinction is the paper’s useful contribution. It separates informational volume from interaction coupling.
Agency lowers predictability because freedom has a cost
The framework then introduces a bound. For classical systems under Shannon information, $P \leq 0.5$. In the quantum formulation, using quantum mutual information, maximally entangled systems can reach $P = 1$. In practical classical agency, the attainable value is lower than the classical ceiling because adding action adds degrees of freedom.
This is the key conceptual inversion: action is not free.
A passive deterministic system can, in principle, approach the classical ceiling if the state representation is complete enough and the dynamics are sufficiently deterministic and invertible. Once a system selects actions, it gains freedom, but that freedom introduces asymmetry. Different internal states and actions may lead to similar outcomes. Similar states may lead to different outcomes. The loop becomes harder to predict in both directions.
The paper tracks this with two directional uncertainties:
This is forward predictive uncertainty: given what the agent knew and did, how uncertain is the outcome?
This is backward predictive uncertainty: given the outcome, how ambiguous are the internal state and action that produced it?
Their difference is:
This matters because two systems can have similar overall $P$ while failing for different reasons. High $H_f$ means the world has become opaque to the agent: actions no longer reliably constrain outcomes. High $H_b$ means the agent has become illegible: outcomes no longer reveal which state-action pattern produced them.
That is not philosophical embroidery. In operations, it is the difference between changing the observation pipeline, damping the action policy, widening the context, narrowing the action space, or simply stopping the system before it confidently automates nonsense at scale. A proud milestone for modern enterprise software, but not one to pursue deliberately.
The double pendulum is the calibration test, not the main sales pitch
The first experiment is a deterministic physical baseline: a double pendulum without an action channel. The authors run two batches of 300 simulations across symmetric and asymmetric mass configurations, using a full phase-space representation.
The purpose is calibration. If $P$ is meaningful, a deterministic physical system with a sufficiently complete state representation should approach the classical ceiling and show little directional asymmetry.
That is what the paper reports. The first batch has mean $P = 0.475747126$; the second has mean $P = 0.472919123$. Both are close to the classical ceiling of $0.5$. Forward and backward uncertainties are also nearly identical, with $\Delta H$ centered near zero.
The robustness point is also useful: chaoticity does not destroy this symmetry. The paper uses finite-time Lyapunov exponents to assess chaotic dynamics and reports no degradation of $P$ with increasing chaos. The likely purpose of this test is not to prove that chaos is harmless in every representation. It is to show that chaotic sensitivity is not the same thing as directional loss of predictability.
That distinction matters later. When the RL agents show lower $P$ and nonzero $\Delta H$, the paper can argue that the difference is not simply “complex systems are noisy.” It is the introduction of agency.
| Test | Likely purpose | What it supports | What it does not prove |
|---|---|---|---|
| Double pendulum, two simulation batches | Physical calibration | $P$ approaches the classical ceiling in deterministic non-agentic dynamics | That all real physical systems will be easy to monitor |
| Lyapunov comparison | Robustness/sensitivity check | Chaos does not itself imply directional asymmetry at the chosen representation | That discretization choices never matter |
| HalfCheetah perturbations | Main engineering validation | $P$, $H_f$, $H_b$, and $\Delta H$ detect coupling degradation earlier than reward | That the same thresholds transfer unchanged to all robots |
| LLM dialogue perturbations | Cross-domain extension | Token-statistical coupling metrics can flag conversational drift | That token-frequency $P$ equals semantic understanding |
The double pendulum, then, is not the headline. It is the ruler placed on the table before measuring the agent.
HalfCheetah shows why reward is too late
The main operational result comes from reinforcement learning agents in MuJoCo’s HalfCheetah environment. The authors evaluate SAC and PPO agents with frozen policies, computing $P$, $H_f$, $H_b$, and $\Delta H$ over sliding windows. They then inject eight perturbation types spanning agent-side degradation and environment-side changes: actuator noise, observation noise, external torso force, and gravity changes.
Under normal operation, HalfCheetah shows:
| System | $P$ | $\Delta H$ | Interpretation |
|---|---|---|---|
| Double pendulum | approximately 0.48 | approximately 0 | Passive physics: high bi-predictability, symmetric prediction |
| HalfCheetah baseline | $0.33 \pm 0.02$ | $-0.56 \pm 0.22$ | Agency: reduced bi-predictability, asymmetric prediction |
The lower $P$ is not a defect by itself. It is the informational cost of having an action channel. The negative $\Delta H$ indicates that backward ambiguity exceeds forward uncertainty: outcomes do not uniquely reveal the internal state-action causes that produced them.
Then comes the practical part.
Across 168 perturbation trials, the paper reports that the Information Digital Twin signal—defined as the union of threshold crossings across $P$, $\Delta H$, $H_f$, and $H_b$—detected $89.3 \pm 15.1%$ of perturbations. Reward-based detection found $44.0 \pm 26.1%$. Median detection latency was 42 windows for the IDT union versus 184 windows for reward.
This is not a small gap. It is the sort of gap that turns “the model degraded” from a post-mortem paragraph into an operational signal.
The mechanism explains the advantage. Reward is an accumulated task-performance measure. It often changes only after degradation has already affected enough transitions to show up in returns. $P$ and $\Delta H$ watch the interaction structure at the transition level. They can move when the agent’s grip weakens, even if the reward has not yet admitted the problem.
For business readers, this is the paper’s most concrete value proposition: coupling telemetry may detect silent degradation before outcome telemetry does.
Not “replace reward.” That would be silly. Reward still tells you whether the objective is being met. But reward does not necessarily tell you whether the agent still understands the loop it is acting inside. That is a different variable.
The LLM experiment extends the idea, with a narrower claim than the headline might suggest
The paper then tests whether the same framework generalizes beyond physical control. A Llama 3.1 8B student model interacts with teacher models across multi-turn conversations. The teacher models include Claude, ChatGPT, and Gemini. The study covers 34 unique test–teacher–condition combinations and 4,574 turns. The dialogue loop is mapped as:
- $S$: accumulated context;
- $A$: the student model’s current response;
- $S’$: the teacher’s subsequent prompt.
The metrics are computed from token-frequency distributions. That detail matters. The LLM experiment is not measuring “understanding” in the human sense, and it is not claiming that token-frequency entropy is a complete model of conversation quality. It is testing whether a lightweight structural coupling signal tracks conversational disruption.
The paper compares $P$ and $\Delta H$ against embedding-based cosine similarity and LLM-as-judge scores. $P$ correlates significantly with structural similarity in 29 of 34 conditions, or 85%. It correlates with judge-based semantic scores in 15 of 34 conditions, or 44%. $\Delta H$ shows a similar pattern: 76% for structural similarity and 47% for semantic scores.
That separation is useful. It says the metric is mostly tracking interaction structure, not semantic correctness. This is both a strength and a boundary.
The perturbation tests are more dramatic. Contradictions, topic shifts, and non-sequiturs are injected after a baseline period. Using token statistics, $P$ and $\Delta H$ reportedly achieve 100% detection across teacher models and perturbation types, matching semantic judges in sensitivity while avoiding heavier semantic evaluation.
The business implication is not that token entropy can replace human evaluation, embedding models, or LLM judges. It is that a cheap structural signal may be useful as an always-on monitoring layer. Semantic judges are too expensive and too slow to run on every turn of every deployed agent conversation. A structural coupling metric may run closer to the stream.
That is a more modest claim. It is also the more useful one.
The paper’s agency-versus-intelligence line is operational, not metaphysical
The title-level claim is provocative: current AI has agency but not intelligence. That could easily become another sterile definitional argument, the kind that makes AI panels feel longer than they are.
The paper avoids some of that by defining the terms operationally.
Agency requires three conditions:
| Agency condition | Meaning | ||
|---|---|---|---|
| Choice: $H(A \mid S) > 0$ | The action is not fully determined by the available state | ||
| Effect: $MI(A; S’ \mid S) > 0$ | The action changes what happens next | ||
| Predictive asymmetry: $ | \Delta H | > 0$ | The system introduces directional intervention into the loop |
By this definition, RL agents and LLM agents are agentic. They select among possible actions or responses. Those selections affect subsequent states or dialogue context. Their interaction loops show asymmetry.
Intelligence then requires more:
| Intelligence condition | Meaning |
|---|---|
| Learning | Increase interaction predictability toward objectives |
| Self-monitoring | Measure and regulate $P$ over time |
| Adaptation | Reshape the state, action, or outcome spaces when coupling degrades |
This is where the paper draws the gap. Current AI systems can learn within a designer-defined interface. They can act within a designer-defined action space. They can optimize toward a designer-defined objective. But they typically do not compute their own coupling quality from their own interaction stream. Nor can they reorganize what they observe, what they can do, or what outcome space they are operating within when that coupling breaks.
That is the “not intelligence yet” claim. It does not say today’s models are useless. It says they lack a regulatory layer that would let them notice and repair degradation in the loop itself.
For operators, this is the more valuable framing. The practical question is not whether a model deserves the word intelligence. The practical question is whether a deployed agent can tell when its interaction structure is becoming unreliable before the business process discovers it the usual way: through customer complaints, compliance incidents, broken workflows, or a suspiciously cheerful incident report.
The Information Digital Twin is a sidecar for coupling, not a smarter model
The paper’s architectural proposal is the Information Digital Twin, or IDT. The name is slightly grand, but the idea is straightforward: run a parallel monitoring layer beside the agent-environment loop.
The IDT receives copies of observations, actions, and outcomes. It computes $P$ and $\Delta H$ in real time. A controller detects deviations from a baseline. If coupling degrades, modulation mechanisms can adjust the interface: dampening actions, filtering inputs, reducing dimensionality, gating context, or adjusting generation parameters.
The authors compare this to thalamocortical regulation in biological systems, where signal statistics can be monitored and modulated without directly interpreting semantic content. The paper is careful not to claim the thalamus implements this exact metric. It uses the biological analogy as an architectural precedent: a copy-based regulatory loop can coexist with task behavior.
The important engineering distinction is this:
| Layer | Role |
|---|---|
| Agent | Optimizes task behavior |
| IDT | Monitors coupling integrity |
| Controller | Detects statistical deviation |
| Modulation pathway | Adjusts observation or action interface |
That separation is clean. It avoids stuffing every safety and reliability function into the model itself. Instead of asking the agent to be its own auditor, mechanic, and therapist, the IDT gives the system an externalized “first-person” coupling monitor. Finally, a division of labor. Civilization advances.
However, the modulation pathways are not fully validated in the paper. The reported experiments validate monitoring and detection. The architecture specifies how modulation could work, but the control laws mapping a particular $P$ or $\Delta H$ deviation to a particular intervention remain domain-specific future work.
That boundary matters. The paper demonstrates a promising alarm and diagnostic signal. It does not deliver a universal self-healing agent.
The business value is earlier diagnosis, not a magic definition of intelligence
For AI builders, the paper’s most useful idea is not the definitional fight over intelligence. It is the operational separation between task performance and interaction coupling.
A deployed agent can fail in at least four different ways:
| Failure mode | What ordinary monitoring may see | What coupling telemetry adds |
|---|---|---|
| Reward degradation | The objective score declines | Confirms whether the loop degraded before the score moved |
| Input drift | Incoming data distribution changes | Tests whether the agent-outcome relation actually lost coherence |
| Semantic drift | Conversation becomes less relevant | Adds a cheap structural signal before expensive judging |
| Action illegibility | Outcomes stop revealing agent intent | Uses $H_b$ to identify backward ambiguity |
This is especially relevant for autonomous workflows. In a human-operated process, workers often notice when the world stops behaving normally. In an automated process, the system may continue executing perfectly within its local policy while the environment quietly stops matching the assumptions under which that policy works.
That is not hypothetical. Any business that deploys agents into pricing, trading, routing, procurement, customer support, compliance triage, or industrial control will eventually face the same ugly truth: a model can remain locally competent while becoming globally miscoupled.
The paper suggests three possible uses:
-
Always-on agent telemetry. Track $P$, $H_f$, $H_b$, and $\Delta H$ over the stream of state-action-outcome data, rather than waiting for reward or task metrics to fail.
-
Failure attribution. Use directional uncertainty to distinguish whether degradation comes from outcome unpredictability, agent ambiguity, or both.
-
Triggering bounded intervention. Use coupling deviations to trigger conservative controls: slow down action cadence, narrow the action space, request human review, reset context, or switch to a safer policy.
These are Cognaptus inferences from the paper’s results, not direct product claims from the authors. The paper shows that the signals are computable and useful in specific physical, RL, and LLM settings. Turning that into production-grade reliability infrastructure requires engineering choices around representation, discretization, thresholding, latency, and intervention design.
Still, the direction is attractive because it is model-agnostic. The IDT does not require access to model internals, reward shaping, or semantic content in the reported experiments. In enterprise environments where agents are increasingly assembled from external models, APIs, tools, and workflow glue, that matters.
A monitoring layer that works from interaction traces is easier to deploy than one that requires opening the model’s skull. Especially when the skull belongs to a vendor.
The limits are not footnotes; they define where the idea can travel
There are three boundaries to keep clear.
First, the formal results assume discrete variables with finite entropies. For continuous systems, the paper computes $P$ after discretization at a fixed resolution. That means the metric depends on representation choices. The double pendulum robustness tests help, and the RL protocol uses structured dimensionality reduction, but production systems would still need careful calibration.
Second, the LLM experiment uses token-frequency distributions. That makes the method lightweight, but it also limits interpretation. A structural coupling signal can flag drift without understanding meaning. That is useful for monitoring. It is not a substitute for semantic evaluation when the business question is whether the answer is correct, compliant, or strategically appropriate.
Third, the IDT’s modulation layer is more architectural than fully demonstrated. The paper validates detection more strongly than recovery. It argues that coupling metrics can supply the regulatory signal that adaptation would need. It does not prove that any particular modulation policy will restore performance across domains.
These limitations do not weaken the core contribution. They prevent over-selling it. A rare and welcome service.
The real shift is from model intelligence to interaction architecture
The AI industry has spent years treating intelligence as something located inside the model: parameters, representations, reasoning traces, benchmarks, tool calls, memory, planning. This paper shifts attention to the interaction boundary.
That is the right place to look for deployed reliability.
A model can be powerful and still poorly coupled to its operating environment. An agent can act and learn without monitoring whether its action loop remains coherent. A system can optimize an objective while losing the informational grip that made the objective meaningful in the first place.
The paper’s cleanest sentence could be reconstructed like this: agency is the ability to act on predictions; intelligence requires monitoring whether those predictions still work and adapting the interface when they do not.
That is a useful distinction because it maps to architecture.
If your AI system has only a model and an objective, it may be agentic. If it also has a self-monitoring coupling layer and a way to adapt what it observes and does, it starts to resemble something more robust.
The current generation of AI systems can produce impressive action. What they generally lack is the machinery to notice when action has become structurally detached from the world. They can keep moving. They can keep answering. They can keep optimizing.
So can a Roomba with one wheel off the floor.
The gap between agency and intelligence is not that today’s systems fail to act. It is that they do not yet regulate the informational cost of acting.
And that distinction is no longer only philosophical. It is becoming measurable.
Cognaptus: Automate the Present, Incubate the Future.
-
Wael Hafez, Chenan Wei, Rodrigo Pena, Amir Nazeri, and Cameron Reid, “A Mathematical Theory of Agency and Intelligence,” arXiv:2602.22519, 2026. https://arxiv.org/pdf/2602.22519 ↩︎