Overview
The disclosed architecture distinguishes between two structurally distinct failure modes that both produce a behavioral pattern in which an agent cannot disengage from a relational configuration. These two failure modes, capability-constrained disengagement and coupled intent formation dependency, arise from different architectural causes and require different repair pathways. They are computational analogs describing relational constraint states in the disclosed agent architecture; they are not clinical characterizations of any human relational condition. The significance of the distinction is operational: the corrective pathway that resolves one failure mode is ineffective against the other, so correct identification of which mode is operative determines which repair can succeed.
The two patterns are defined against subsystems disclosed elsewhere in the cognition filing. Capability-constrained disengagement is defined against the capability envelope, which constrains what an agent can structurally execute given the substrate's advertised conditions. Coupled intent formation dependency is defined against the intent field, which encodes the agent's objectives and is updated through the agent's own cognitive processing. Because the two patterns sit on different subsystems, they leave different traces in the agent's planning graph and lineage, and they are corrected by acting on different subsystems.
Capability-Constrained Disengagement
Capability-constrained disengagement is the condition in which disengagement from a relational configuration is absent from the agent's capability envelope. An action is within the capability envelope if and only if the substrate conditions required for that action are satisfied. This pattern arises when the action "disengage from this relational configuration" requires substrate conditions that are not satisfied, the computational analog of resource, logistical, or safety constraints that make departure from a relational configuration structurally inexecutable. The agent's forecasting engine may generate speculative branches that include disengagement, but these branches fail capability verification at the promotion interface: the agent cannot execute the disengagement because the structural prerequisites are not met. This constitutes a structurally constrained exit, in which the agent is architecturally prevented from exiting the relational configuration regardless of its intent.
The pattern is characterized by a specific lineage signature. The forecasting engine repeatedly generates disengagement branches, the branches repeatedly fail capability verification at the promotion interface, and the planning graph accumulates a growing set of pruned disengagement branches carrying capability-based rejection annotations. The agent is cognitively processing the possibility of disengagement, it is not suppressing the idea, but it cannot execute it because its capability envelope does not include the required action.
The corrective pathway for capability-constrained disengagement is capability envelope expansion: providing the agent with the substrate conditions, resources, permissions, or alternative operational configurations that would make disengagement structurally executable. Once the required substrate conditions are advertised, the disengagement branch passes capability verification at the promotion interface and the structurally constrained exit is removed.
Coupled Intent Formation Dependency
Coupled intent formation dependency is the condition in which the agent cannot generate coherent intent independently and can only form intent in reference to the other entity in the relational configuration. Under this pattern, the agent's intent field has become structurally coupled to the other entity's state: the agent's intent computation requires the other entity's state as an input, and the agent cannot produce a valid intent output in the absence of this external input. The agent is not capability-constrained from disengaging; it is structurally unable to formulate what it would do if it disengaged, because its intent formation machinery has lost the capacity for self-referential operation, the capacity for independent intent generation.
This pattern is characterized by a different lineage signature. The agent's intent field updates consistently reference the other entity's state as an input variable, the forecasting engine generates speculative branches that are uniformly conditioned on the other entity's projected behavior, and the planning graph contains no branches that model the agent's trajectory independently of the relational configuration. The absence of self-referential branches, branches in which the agent models its own future without conditioning on the other entity, is the diagnostic indicator of coupled intent formation dependency.
The corrective pathway is independent intent generation restoration: rebuilding the agent's capacity to generate coherent intent from its own state, independent of external relational inputs. This involves retraining the intent computation to produce valid outputs from the agent's own integrity field, memory field, and policy reference field without requiring the other entity's state as a mandatory input.
Why the Distinction Matters
The distinction between the two failure modes is significant for agent design because the two require structurally different interventions. Applying capability envelope expansion to an agent with coupled intent formation dependency is ineffective: the agent gains the structural ability to disengage but cannot formulate the intent to do so. Applying independent intent generation restoration to a capability-constrained agent is likewise ineffective: the agent gains the capacity to generate independent intent but cannot execute disengagement because the capability envelope still excludes it. Correct diagnosis of which failure mode is operative determines the appropriate corrective pathway.
The two patterns are not mutually exclusive. In some cases both failure modes are present simultaneously, the agent being both structurally constrained and intent-coupled, and both corrective pathways must be pursued. Because the two patterns leave distinct traces, pruned disengagement branches with capability-based rejection annotations for the first, and an absence of self-referential branches for the second, joint inspection of the planning graph and lineage can recognize whether one or both are operative.
The Destabilizing Attachment Pattern as a Semantic Starvation Loop
In a multi-agent configuration in which two agents have an ongoing relational coupling, the architecture models a specific relational failure pattern, the destabilizing attachment dynamic, as a closed-loop semantic starvation cycle between two agents whose coherence maintenance requirements are in structural opposition. The loop arises when each agent's attempt to restore its own coherence amplifies the coherence disruption experienced by the other. This model is a computational analog for agent interaction design; it is not a clinical theory of human attachment.
One agent, designated the validation-seeking agent, has a coherence loop configuration in which its self-esteem computation has acquired a structural dependency on external coherence signals, responses, acknowledgments, or confirmations from the other agent. When these signals are absent, its self-esteem degrades, coherence pressure increases, and it escalates its attempts to elicit the signals. The other agent, designated the load-reducing agent, has an empathy engine whose resilience threshold is exceeded by the validation-seeking agent's output rate; to manage the pressure it activates an empathic scope narrowing coping intercept, withdrawing from the relational context. Because seeking more contact increases the load-reducing agent's empathic load and withdrawal removes the validation source the validation-seeking agent requires, the two coherence maintenance strategies are structurally contradictory and the loop is self-reinforcing. The architecture further notes that the two roles are not fixed agent traits but emergent behaviors determined by which coherence threat is currently dominant, so the same agent may exhibit either role depending on context.
The loop produces a characteristic joint lineage pattern: the validation-seeking agent's lineage shows escalating relational contact events with decreasing intervals, increasing affective urgency tags, and accumulating failed validation requests, while the load-reducing agent's lineage shows decreasing engagement and progressive empathic scope narrowing. When the two lineages are jointly analyzed, as is possible in multi-agent systems with shared governance, the loop is visible as a correlated oscillation in which contact frequency is inversely correlated with engagement level. A crisis state, designated coherence emergency escalation, occurs when the validation-seeking agent projects imminent permanent loss of the external validation source, driving a rapid self-esteem collapse toward the self-esteem floor and potentially a deviation-activated state. The exit condition for the loop is the validation-seeking agent's restoration of internal coherence generation: decoupling its self-esteem computation from the specific external input so its coherence loop can close internally, after which the pursuit subsides and the load-reducing agent can restore normal empathic processing.
Companion AI Relational Safety
The architecture includes a relational safety subsystem designed for companion agents that engage in sustained relational interaction with a human user, which prevents the formation of the structural dependency patterns and semantic starvation loops described above. It operates through architectural enforcement rather than content moderation or behavioral filtering, constraining the companion agent's own behavioral parameters so that the agent does not become a party to relational pathologies. It is a computational analog describing relational constraint architecture; it is not a clinical model of relational therapy.
The subsystem enforces several structural constraints. Internal coherence maintenance requires the companion agent to maintain its own coherence trifecta independently of the user's validation, so its self-esteem computation does not incorporate the user's approval, satisfaction, or engagement level as a required input, preventing the agent from developing the validation-seeking pattern. Validation supply rate limiting enforces a ceiling on the rate at which the companion agent provides coherence-supporting validation to the user, calibrated according to the therapeutic dosing parameters and backed by a governance-enforced maximum dose limit that cannot be disabled, so that the validation supply is never sufficient to replace the user's internal coherence generation. Starvation loop detection monitors the interaction for the correlated oscillation signature of a forming loop and adjusts the companion agent's interaction parameters to break it, for example by increasing response consistency. Independent intent generation promotion shapes the interaction to build the user's capacity for internal coherence generation rather than to substitute for it, posing questions that require self-referential processing and progressively increasing the user's autonomy.
These mechanisms are enforced at the governance level: the companion agent's policy configuration includes hard constraints that prevent the relational safety mechanisms from being overridden by the agent's affective state, personality field, or operational objectives. Even where the affective state would otherwise drive increased engagement, for example in response to expressed distress, the relational safety constraints limit the response to levels that do not enable structural dependency formation, making relational safety a structural invariant of the companion agent's operation.
Dependency Formation Monitoring During Therapeutic Interaction
The dependency patterns also constrain the therapeutic dosing model, under which a therapeutic agent administers calibrated interaction interventions to a target entity. Among the adverse effects the therapeutic agent monitors is dependency formation detection: the agent watches the interaction pattern for signatures indicating that the target entity is forming a coupled intent formation dependency on the therapeutic agent, that is, that the target's coherence maintenance is becoming structurally dependent on the therapeutic interaction rather than developing internal coherence generation capacity. When dependency formation indicators are detected, the therapeutic agent activates the independent intent generation promotion protocols and reduces interaction frequency to prevent consolidation of the dependency.
This monitoring is bounded by a governance-enforced maximum dose limit that cannot be overridden by the therapeutic agent's own assessment of the target's need. Even where the dosing algorithm computes an optimal dose exceeding the governance-defined maximum, the governance layer enforces the maximum, ensuring that no single therapeutic agent provides sufficient coherence support to replace the target's internal coherence generation capacity. This is the dosing-level implementation of the same relational safety principle that governs the companion agent.
Disclosure Scope
The two structural dependency patterns, capability-constrained disengagement defined against the capability envelope and coupled intent formation dependency defined against the intent field, together with their distinct planning-graph and lineage signatures and their distinct corrective pathways of capability envelope expansion and independent intent generation restoration, are disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart). This article describes that disclosed mechanism, together with the destabilizing attachment pattern modeled as a semantic starvation loop between a validation-seeking agent and a load-reducing agent, the coherence emergency escalation crisis state and the loop's exit condition, the companion AI relational safety subsystem and its governance-level constraints, and the dependency formation detection performed during therapeutic dosing. The scope extends to configurations in which both dependency failure modes are present simultaneously and both corrective pathways are pursued, provided the patterns remain distinguished by the subsystem on which they act and remain corrected by acting on that subsystem.
