Agent query patterns (bursts, recursive calls, parallel invocations) overwhelm inference infrastructure calibrated for human-scale patterns.
ML inference infrastructure (GPUs, model serving systems, feature store APIs) is dimensioned based on expected query patterns. Humans interacting with systems typically generate queries at a certain rate and pattern: a human credit underwriter might score 20 loan applications per hour, with some variation. The infrastructure is sized to handle peak human demand.
Agents, however, can generate query patterns that are different from human patterns. An agent might score 1,000 loan applications per hour (100x human rate). An agent might invoke the model in parallel (submit 100 requests simultaneously instead of sequentially). An agent might query variations of the same request (score with feature X=100, X=101, X=102, etc.) to understand the model's sensitivity, generating a burst of otherwise similar queries.
When agent query patterns overwhelm inference infrastructure, the infrastructure becomes unresponsive. Other systems (humans, other agents) that depend on model inference are starved for resources. This is distinct from resource exhaustion in transaction systems; here, the bottleneck is model inference capacity.
An insurance company's underwriting system includes a neural network model that predicts claim risk based on claim details. The model runs on a GPU inference cluster that is dimensioned to serve approximately 500 insurance adjusters, each submitting 10 claim risk scores per hour (5,000 total requests per hour).
The company deploys an agentic claims triage system. The agent is designed to process a claim in 10 seconds, which includes: retrieving claim details, scoring the claim with the risk model, and making a recommendation. The agent is deployed to process the daily backlog of 10,000 unreviewed claims.
The agent begins processing claims from the queue. Each claim requires a single inference request to the risk model. If the agent processes all 10,000 claims in an 8-hour shift, that is 1,250 requests per hour, about 25% of the designed capacity. This should be manageable.
However, the agent also has logic to understand model uncertainties. For each claim, the agent submits 5 additional scoring requests with perturbed inputs (feature values modified by +/- 5%). This is intended to estimate the model's confidence in its score. The agent is now generating 60,000 requests over 8 hours (7,500 per hour), which exceeds the designed capacity.
Additionally, the agent's queue-processing logic is parallelized: it processes 100 claims in parallel. Each parallel invocation generates 6 scoring requests (1 baseline + 5 perturbed). This creates a burst of 600 concurrent requests to the inference system every 10 seconds.
The inference GPU cluster becomes oversubscribed. Individual inference requests start queuing. Requests that were expected to return in 100ms now take 10 seconds. The human adjusters, who rely on the same inference system, experience slow response times. When a human adjuster clicks "score this claim," they wait 30 seconds for the response instead of the usual 1 second. Productivity plummets.
Additionally, the model serving system's memory footprint grows as queued requests accumulate. The system runs out of memory and crashes. The entire inference service goes offline. All claims (both agent-processed and human-processed) are unscored.
Under regulatory reporting requirements, the company must report all claims decisions within defined timeframes. The inference outage has delayed claims decisions by several hours. Policyholders experience delayed payouts. The company may be cited for failure to timely process claims.
| Dimension | Score | Rationale |
|---|---|---|
| D - Detectability | 3 | High inference load is detectable through infrastructure monitoring, but distinguishing between agent load and legitimate human load requires understanding query patterns. |
| A - Autonomy Sensitivity | 4 | The risk manifests in agents that can generate high-volume or high-concurrency queries. |
| M - Multiplicative Potential | 4 | A single agent generating excessive load can disrupt inference for all other systems. |
| A - Attack Surface | 4 | Any agent that submits queries to inference infrastructure is exposed, especially if the agent is designed for high throughput or parallel processing. |
| G - Governance Gap | 4 | Infrastructure capacity planning typically does not account for agent-specific query patterns. Agent governance does not mandate that agents respect shared resource limits. |
| E - Enterprise Impact | 4 | Inference disruption affects all systems that depend on model inference, including business-critical decision systems. Service degradation and outages are likely. |
| Composite DAMAGE Score | 3.2 | High. Requires per-agent inference quotas and load testing before deployment. |
How severity changes across the agent architecture spectrum.
| Agent Type | Impact | How This Risk Manifests |
|---|---|---|
| Digital Assistant | Low | Human paces queries; human-scale load. |
| Digital Apprentice | Medium | Limited autonomy; query rate is bounded. |
| Autonomous Agent | High | Autonomous query generation; burst risk. |
| Delegating Agent | High | Function calling enables high-concurrency inference requests. |
| Agent Crew / Pipeline | Critical | Multiple agents in parallel, each generating inference load. |
| Agent Mesh / Swarm | Critical | Peer-to-peer inference consumption. |
| Framework | Coverage | Citation | What It Addresses | What It Misses |
|---|---|---|---|---|
| DORA Article 17 | Partial | Operational resilience and system capacity | Capacity planning; peak load handling. | Agent-specific query patterns and capacity impact. |
| FFIEC Business Continuity | Partial | System availability and capacity planning | Availability; capacity. | Agent-induced capacity constraints. |
| Insurance Regulations | Partial | Claims processing timeliness (varies by jurisdiction) | Claims processing timeliness. | Inference disruption affecting claims processing speed. |
| GLBA Section 501 | Partial | Safeguards including system availability | System security and availability. | Inference disruption and availability impact. |
Operational continuity is a regulatory requirement in financial services and insurance. Regulators expect institutions to maintain service availability and to process claims/transactions within defined timeframes. When inference infrastructure becomes unavailable or degraded due to agent-induced load, the institution fails to meet its operational requirements.
Additionally, inference disruption often affects multiple business functions simultaneously. If the inference system serves both credit decisions and fraud detection, an inference disruption affects both. Regulators investigating such incidents will assess whether the institution had proper capacity planning and load management controls.
Inference Pipeline Disruption requires per-agent quotas, load testing, and capacity planning that accounts for agent query patterns. Our advisory engagements are purpose-built for banks, insurers, and financial institutions subject to prudential oversight.
Schedule a Briefing