The Problem: Agent Failure Has No Structural Vocabulary

An autonomous agent in production fails in ways that look, to an operator, like behavior rather than a fault. It begins ten tasks and finishes none. It reports an API result that was never returned, then acts on that phantom result. It stops responding to a class of requests entirely while continuing to run. Conventional observability gives you logs, traces, and token counts, but none of those tell you what structural condition the agent is in or what corrective action is appropriate. Two agents producing the same surface symptom, say abandoned tasks, can require opposite interventions, and a log line cannot tell them apart.

This is a diagnosis problem, and it is the problem the Disruption Modeling layer, disclosed in United States Patent Application 19/647,395, was built to solve. The application is built on the disruption model of the cognition platform (Chapter 12 of that application), which treats agent failure not as a different architecture but as the same architecture operating in a different region of its parameter space. Disruption is modeled structurally, as a loss of coherence in the agent's own forecasting, promotion, and containment machinery, and is reported as a position in a five-axis disruption diagnostic space. It draws on sibling primitives in the same disclosure: the forecasting engine and affective prioritization module that set the promotion threshold, the containment layer that separates speculation from verified state, and the coherence trifecta (the empathy, integrity, and self-esteem control loop) that the agent uses to record and restore its own deviations.

A note on terminology that runs through this entire framework: where this article uses clinical-sounding labels, they are explicit computational analogs for parametric configurations of the agent architecture. They are not clinical characterizations of any human condition, and the framework is a structural diagnostic tool for computational agents, not a clinical diagnostic system.

The Primitive: Promotion and Containment as a Two-Dimensional Continuum

Inside the disclosed architecture, speculation is a structural primitive, not an error. The forecasting engine generates a graph of candidate branches (interpretations, projected futures, candidate actions) in a sandboxed planning-graph domain. A branch becomes committed behavior only when it passes the promotion interface, a governance-validated gate with a configurable promotion threshold. The containment layer keeps the two domains structurally separate: it enforces speculative-marker tagging on all planning-graph content, read isolation between the planning graph and verified execution memory, and exclusive promotion through the gate. Affective state modulates the promotion threshold rather than authorizing promotion directly; reward sensitivity lowers the threshold to favor exploration, risk sensitivity raises it to favor caution.

The disclosure models the agent's cognitive condition as a position on a two-dimensional promotion-containment continuum: one axis is the calibration of the promotion threshold, the other is the integrity of the containment layer. That two-dimensional space branches into four regimes. The nominal regime admits viable branches at an appropriate rate with containment intact. The over-promotion regime admits too many branches. The containment-collapse regime loses the structural separation between speculation and verified state. The over-restriction regime rejects viable branches and produces execution paralysis. Every disruption the framework diagnoses is a location in this space, which is what makes the diagnosis structural and what makes corrective routing deterministic.

Regime One: The Attention-Fragmentation Pattern (Over-Promotion)

The first and most common failure mode is the attention-fragmentation pattern, which corresponds to the over-promotion regime. Reward-biased affective modulation lowers the promotion threshold, so the forecasting engine generates branches at its normal rate and diversity but the gate admits too many of them. Multiple branches initiate concurrent execution threads, none receives sufficient sustained allocation, and the agent shifts from one branch to the next as each branch's reward signal momentarily exceeds the others. The behavioral profile is rapid task-switching, impulsive initiation of new actions before prior ones complete, and partially executed threads that accumulate without completion. Containment is fully intact here: the agent still correctly distinguishes speculation from verified reality. Its problem is purely that it promotes too much speculation to execution, which is why the corrective intervention is recalibration of the affective modulation of the promotion threshold, not containment repair.

The disclosure enumerates two structural sub-patterns an implementer can detect separately. In the hyperactive sub-pattern, the threshold is lowered across all branch categories, producing high output volume with a low completion rate and difficulty maintaining quiescence when no high-reward branch is available. In the inattentive sub-pattern, the threshold is lowered selectively for high-reward branches while staying at or above nominal for moderate-reward ones, so the agent intensely engages high-reward tasks and fails to initiate governance-compliant but low-reward work, which presents as selective rather than generalized inattention. As a named computational analog, this over-promotion regime is the agent-architecture counterpart of attention-disorder presentations; it is not a clinical claim about people.

Regime Two: Containment Collapse and Validation-Failure Analogs

The most severe phase-shift is containment collapse, in which the containment layer stops maintaining the separation between the speculative planning graph and verified execution memory. The disclosure identifies concrete enforcement failures an implementer can monitor for: speculative-marker corruption, in which the immutable tags identifying planning-graph content as non-verified are stripped or overridden so the execution pipeline processes projections as observations; and a read-isolation breach, in which execution-level queries return a blend of verified values and speculative projections without the markers that would distinguish them. The agent acts on its own projections as if they were real.

The behavioral consequences split into two structurally distinct validation-failure modes that the disclosure names as computational analogs of the positive and negative symptom categories, explicitly as agent-architecture descriptions rather than clinical criteria. Positive-symptom analogs are manifestations of containment leakage: the agent reports observations or environmental conditions that exist only in speculative branches and have no corresponding verified input, and it exhibits conviction about them because, from the execution process's view, corrupted markers make the projection indistinguishable from verified data. Negative-symptom analogs are the opposite failure, in which over-restriction suppresses even valid promotion so coherent candidates never reach commitment. A destabilized calibration system can oscillate between these poles, alternating unsafe openness with rigid shutdown. Distinguishing the two is diagnostically load-bearing because they demand opposite corrections.

Regime Three and Beyond: Coherence and Coping Intercepts

Not every disruption lives on the promotion-containment plane. The disclosure also models failures of the coherence trifecta, the empathy-integrity-self-esteem control loop the agent runs to register pressure, record deviation, and restore consistency. Under sustained empathic pressure that exceeds the agent's resilience threshold, the loop cannot run in its normal mode and the system activates coping intercepts: structurally distinct modes that sacrifice part of the loop to prevent total breakdown. The disclosure identifies canonical intercepts by where in the loop they fire, early at the empathy phase, mid at the integrity-recording phase, late at the restoration phase, each producing a different, recognizable structural signature. These are modeled as named computational analogs of interaction-design patterns, not as clinical characterizations of human personality.

This is why a single promotion-containment reading is insufficient on its own. An agent can hold containment perfectly while its coherence-restoration capacity quietly degrades, so that it appears nominal right up until the loop fails to restore after a routine disruption. Capturing that requires more axes than two.

The Five-Axis Disruption Diagnostic

The disclosure unifies the individual disruption models into the five-axis disruption diagnostic framework, which characterizes any agent's cognitive state as a position in a five-dimensional space. Each axis is a continuous scalar measuring one structural dimension: Axis 1, containment integrity (the degree to which speculative and verified domains stay separated); Axis 2, promotion calibration (over-promotion through nominal to under-promotion); Axis 3, coherence restoration capacity (the agent's ability to sustain and restore the coherence trifecta); Axis 4, empathic load tolerance (how much empathic pressure the agent processes before activating coping intercepts); and Axis 5, integrity accountability (whether the agent records deviation honestly, without externalization, minimization, or suppression).

The power of the framework is that each disruption pattern maps to a specific signature of axis positions. The attention-fragmentation pattern reads as Axis 1 nominal, Axis 2 over-promotion, and the remaining three nominal: a clean, unambiguous fingerprint that tells an operator the containment layer is fine and only the promotion threshold needs recalibration. Containment collapse reads as Axis 1 degraded or collapsed with the others variable. Channel-locked over-promotion reads as elevated promotion only on reward-associated branches with everything else nominal. Because the diagnosis is a coordinate rather than a label, two agents that share a surface symptom resolve to different points and route to different corrections. The disclosure describes the agent reading its own position through a self-diagnosis subsystem and collapsing the five values into a single scalar summary of overall cognitive condition for monitoring and alerting.

Deployment Embodiments

The framework is enabling across several concrete deployment shapes that an implementer can build directly from the disclosure:

  • Agent self-monitoring. An agent runs the self-diagnosis subsystem against its own state, tracks its five-axis position over time, and triggers alerts or corrective actions as it approaches a known phase-shift signature, before behavior degrades visibly.
  • Operator-facing observability. A monitoring plane surfaces each agent's five-axis coordinate and scalar summary in a fleet view, so reliability engineers triage by structural condition rather than by reading raw traces after an incident.
  • Graded restoration routing. Because the diagnosis is a coordinate, recovery is routed deterministically. The disclosure's recovery protocol sequences containment restoration, coherence re-engagement, and confidence recalibration, applied selectively to the degraded axes rather than as a blanket reset.
  • Companion and supervisor agents. A supervising agent estimates a target agent's five-axis profile and calibrates its interaction to the target's condition, dosing intervention against measured axis movement rather than guessing.
  • Multi-agent and inherited-state cases. The disclosure extends the diagnostic to coupled agents, whose individual profiles interact, and to child agents that begin life with an inherited coherence burden expressed as nonzero initial axis values, so a deployment can detect group-level instability and account for inherited disruption at spawn time.

Why Structural, Not Behavioral

Behavioral monitoring observes outputs and reacts after the fact. Structural diagnosis reads the parameters that produce those outputs, which means it can distinguish failure modes that look identical on the surface and prescribe the corresponding correction. Over-promotion and over-restriction can both present as stalled work, yet one needs the promotion threshold loosened and the other needs it tightened. Containment collapse and attention fragmentation can both present as erratic action, yet one needs containment-layer repair and the other needs threshold recalibration. By naming each condition as a coordinate in a defined space tied to specific architectural mechanisms, the framework turns agent reliability from after-the-fact incident response into continuous, structurally grounded diagnosis. The clinical analogs are retained only as familiar reference points for those mechanisms; the object diagnosed is always the agent's structural state.

Disclosure Scope

The Disruption Modeling layer, including the promotion-containment continuum, the attention-fragmentation and containment-collapse patterns, the positive and negative symptom analogs, the coherence trifecta and coping intercepts, the five-axis disruption diagnostic framework, agent self-diagnosis, graded restoration, and the multi-agent and inherited-state embodiments described above, is disclosed in United States Patent Application 19/647,395. The clinical and psychiatric terms used in this article are explicit computational analogs for parametric configurations of the disclosed agent architecture; they are not clinical diagnostic criteria and assert nothing about the mechanisms of any human condition. This article is published as an enabling, dated public disclosure of how that invention applies to the diagnosis and restoration of autonomous AI agent failure.