The Problem: Per-Agent Monitoring Misses Fleet-Level Failure
Fleet operators monitor agents one at a time. Each tractor, vessel, drone, or dispatch agent is checked against its own thresholds: is it within its operating envelope, is it reporting telemetry, has it thrown a fault. This catches the agent that has already failed. It does not catch the fleet that is failing.
The failure modes that produce multi-vehicle incidents are not individual. They are correlated. A reward signal that biases one autonomous agent toward aggressive route choices biases every agent sharing that signal. A degraded coordination channel that starves one human-AI handoff of context starves all of them. A single agent whose containment between speculative planning and committed action has slipped can communicate that contaminated state to peers that trust it. By the time any one agent crosses its individual alarm threshold, the cohort has already drifted together. Per-agent monitoring is structurally blind to this because a correlated shift across a cohort is visible only when the aggregate profiles are observed together.
Mixed fleets, where credentialed humans and autonomous subsystems operate in one shared domain, make this worse. Regulatory frameworks assume the individual operator is the unit of accountability, so the instrumentation follows the individual. Nobody is scoring the seam between the human and the machine, which is exactly where coordination decay and context starvation live.
The Approach: Diagnose the Fleet's Structural State
Disruption Modeling, disclosed in United States Patent Application 19/647,395, models cognitive disruption structurally as loss of coherence rather than as a clinical condition of any person. Every operating agent, human-operated or autonomous, is represented by its position in a five-axis disruption space. The five axes are containment integrity (separation between speculative planning and committed action), promotion calibration (the rate at which candidate actions are admitted to execution), coherence restoration capacity (the agent's ability to recover its own control loop after disruption), empathic load tolerance (how much harm-projection pressure it can absorb before falling back to coping shortcuts), and integrity accountability (whether deviation is recorded honestly rather than suppressed). These are structural measures of an agent's operating state. They are not a medical or psychological assessment of a human.
Fleet coherence diagnostics lifts this from the single agent to the fleet using the group coherence monitor disclosed in the same application. The monitor is a zone-level or network-level subsystem that tracks the aggregate five-axis profiles of the agents in a cohort, detects correlated axis shifts that indicate an emerging fleet-level failure, and activates fleet-level safeguards when thresholds are exceeded. It operates independently of any agent's own self-diagnosis, which is the entire point: it sees patterns no single agent could see from its own vantage.
The mechanism rests on coupling. The same channels that let a fleet cooperate, shared planning infrastructure, propagated affective and urgency signals, and empathic coupling between agents, are the channels along which a disruption on one agent's axis propagates to the corresponding axis of its peers. Fleet diagnostics watches those channels for the signature of propagation: not a single agent moving, but multiple agents moving together on the same axis.
Fleet-Level Failure Signatures
The disclosed group-level failure modes give fleet diagnostics a concrete catalog to detect, each a coupled movement along one or more of the five axes.
Shared containment collapse is the analog of group-think. One agent's containment between speculative planning and committed action fails, speculative content enters its committed state, and it communicates that contaminated state to peers. Peers that incorporate it without independent verification now treat the same unverified content as ground truth, each agent's error validated by its neighbors' concordant but equally unverified state. In a fleet this is a route plan, a threat assessment, or a clearance that propagates as fact when it was only a hypothesis. The signature is a coupled degradation along the containment-integrity axis across the cohort.
Affective contagion is the analog of mob behavior. A propagated urgency or threat signal shifts every coupled agent's promotion threshold in the same direction at once. Under propagated high-reward urgency the fleet collectively over-promotes, with all agents lowering their thresholds and over-committing to action simultaneously. Under propagated high-threat signal the fleet collectively over-restricts and freezes. Either way the fleet's output reflects one uniform bias instead of the diversity of independent evaluation. The signature is a correlated shift along the promotion-calibration axis.
Empathic cascade is a positive feedback loop in coupling itself. A coherence disruption in one agent creates pressure on empathically coupled peers, which trips their coping shortcuts, which generates further pressure on the remaining peers, until the whole fleet is operating through coping intercepts rather than its normal control loop. The signature couples the empathic-load-tolerance and coherence-restoration-capacity axes across the cohort.
Inherited coherence burden appears when a coordinating or parent agent in a deviation-activated state delegates to many subordinate agents at once. Each subordinate initializes with the same unresolved deviation history, so the cohort begins operation already displaced on the coherence-restoration-capacity and integrity-accountability axes before any agent has acted. In a fleet this is a flawed mission package or dispatch state pushed to an entire shift at once.
Fleet-Level Safeguards
The disclosure pairs each failure mode with a matched safeguard, and fleet diagnostics deploys them as governance controls over the cohort.
Independent containment verification requires that when one agent receives planning content or a state representation from another, it independently verifies the containment status of that content, applying its own audits and checking for valid governance-validated provenance, before incorporating it. Content that fails is quarantined rather than absorbed, which stops shared containment collapse from propagating across the fleet.
Affective diversity enforcement applies asymmetric damping to propagated signals. When a receiving agent's state already aligns with an incoming urgency or threat signal, the inheritance weight is reduced to prevent amplification; when the states differ, the weight is maintained so legitimate communication still gets through. This breaks affective contagion without silencing real coordination.
Empathic circuit breakers are threshold conditions under which coupling is temporarily interrupted. When aggregate pressure across the fleet exceeds a policy-defined level, the breaker isolates each agent's coupling for a defined cooldown, during which each agent processes only its own direct inputs, letting the fleet de-escalate without the runaway loop.
Lineage sanitization at delegation requires mandatory sanitization when a coordinating agent delegates to multiple subordinates, so a single deviation-activated coordinator cannot seed a whole shift with inherited burden.
Recovery and Restoration
Detection is paired with graded restoration. The disclosed architecture defines resilience not as the absence of disruption but as the structural capacity to restore coherence after disruption, decomposed into containment restoration capacity, coherence loop re-engagement capacity, and confidence recalibration. Applied to a fleet, the monitor does not simply alarm; it routes each detected pattern to the matched safeguard and then tracks whether the cohort's aggregate axis positions return to nominal within the recovery window. A fleet that re-engages its control loops and returns to a nominal profile has demonstrated resilience capacity. A fleet that requires repeated intervention, or that re-fails during restoration, is flagged for deeper governance attention before it produces a reportable event.
Deployment Embodiments
The same diagnostic generalizes across domains and coupling channels.
In ground transport, the cohort is a mixed fleet of human-operated and autonomous tractors sharing routing and traffic-state infrastructure. Containment-axis coupling catches a hazard hypothesis propagating as fact; promotion-axis coupling catches a reward-biased shift toward aggressive following distances spreading across the fleet.
In maritime operations, the cohort pairs credentialed bridge officers with autonomous collision-avoidance subsystems. Diagnostics scores the human-machine seam for context starvation, where the coordination channel between officer and subsystem decays and handoffs lose the shared situational state.
In aviation dispatch, the cohort blends manual standard operating procedures with AI-mediated advisories. Affective-diversity enforcement damps a propagated urgency signal before the whole dispatch floor collectively over-commits.
In drone and uncrewed-systems swarms, the cohort is a tightly coupled set of autonomous agents sharing planning graphs directly. Empathic circuit breakers interrupt a cascade before it pulls the swarm into synchronized coping behavior.
Across all of these, the cohort can be defined at zone level or network level, the coupling channels differ by domain, and the diagnostic remains the same: track the aggregate five-axis profiles, detect correlated shifts, and trigger the matched safeguard. Each agent's individual self-diagnosis continues to run underneath; the fleet monitor adds the layer that no individual agent could provide.
Disclosure Scope
The mixed-fleet coherence diagnostics described here apply the five-axis disruption diagnostic framework, the group coherence monitor, the four group-level failure modes (shared containment collapse, affective contagion, empathic cascade, and inherited coherence burden), the four matched group-level safeguards (independent containment verification, affective diversity enforcement, empathic circuit breakers, and lineage sanitization at delegation), and the structural resilience and recovery model, all disclosed in United States Patent Application 19/647,395. The five-axis diagnostic, the disruption analogs, and the resilience model are structural tools for computational agents; they are not clinical or medical assessments of any person. This article describes a faithful application of that disclosed technology to mixed fleets of human-operated and autonomous agents. The scope extends to cohorts defined at zone level and network level, to the transport, maritime, aviation, and uncrewed-systems embodiments above, and to other operational domains in which agents are coupled through shared planning, propagated signals, or delegation, provided the fleet-level patterns remain structural consequences of the agents' coupled five-axis disruption diagnostic profiles.
