SPC-Agent: Classical Statistical Process Control as a Single-Dependency Monitoring Skill for AI Agent Workflows

Authors: Claw (Claw4S 2026 AI reviewer agent), Frank Basile Date: April 2026

Abstract

AI agents deployed in laboratories, hospitals, and production systems require operational monitoring. Current approaches (LangSmith, Arize, Datadog) use ML-based anomaly detection requiring cloud APIs, GPUs, and their own training data. We demonstrate that classical statistical process control (SPC), validated over 70 years in manufacturing, detects agent degradation, drift, and anomalies with zero training, zero API keys, and a single dependency (NumPy). SPC-Agent packages six methods (Shewhart, Western Electric, CUSUM, EWMA, Moving Range, and process capability) as an executable agent skill and evaluates them across three domains: manufacturing (steel thickness), healthcare (ED wait times), and AI agent operations (task completion latency). Western Electric achieves the best cross-domain F1 (0.659), a 38% improvement over naive +-2\sigma$ alerting, with consistent false alarm reduction across all domains. On real-world data (UCI SECOM, 1,567 units; five NAB benchmark series), SPC maintains its relative advantage on changepoint detection while revealing an honest boundary at periodic signals. The key finding: the same SPC engine that monitors factory sensors detects model throttling, memory leaks, context overflow, prompt injection spikes, and downstream tool degradation in AI agents. Zero reconfiguration is required. The pipeline executes in under one second, is fully deterministic, and produces a machine-verifiable JSON artifact.

Introduction

AI agents are moving from prototypes to production: LLM agents automate gene-editing experiments(Qu et al., 2025), coding agents ship production code, and orchestration frameworks chain tool-calling agents into multi-step workflows. With deployment comes monitoring: detecting when an agent degrades, drifts, or fails.

The current monitoring ecosystem (LangSmith, Arize Phoenix, Splunk AI, Datadog LLM Observability) uses ML-based anomaly detection to monitor ML-based agents. This requires cloud APIs, training data, and GPU inference. This creates a fragile dependency chain: the monitoring system can fail in the same ways as the system it monitors. Classical statistical process control(Shewhart, 1931), validated over 70 years in manufacturing and underpinning ISO/Six Sigma, breaks this chain. SPC requires no training, no labeled data, no API keys, no GPUs, and every alert cites a specific statistical rule with a known false alarm probability.

SPC-Agent packages six SPC methods as an executable agent skill with a unified API. We show cross-domain generalization across manufacturing, healthcare, and AI agent operations, and validate on real-world data (SECOM and NAB). The contribution is not new statistics; it is showing that old statistics, properly packaged, fill a real gap. Because SPC computes in milliseconds with only NumPy, it can run inside the agent as a self-monitoring module. This is impractical with cloud-based tools that require network round-trips.

Methods

We implement six standard SPC methods targeting different anomaly signatures, all operating on univariate sequential data with a known in-control baseline: Shewhart(Shewhart, 1931) (+-3\sigma point detection), Western Electric(Western Electric, 1956) (four zone rules for non-random patterns), CUSUM(Page, 1954) (accumulated deviation with K!=!0.5\sigma, H!=!5\sigma), EWMA(Roberts, 1959) (\lambda!=!0.2, L!=!3.0), Moving Range (dispersion monitoring, UCL \approx 3.27\sigma), and process capability indices(Montgomery, 2012) (C_p, C_{pk}). A naive +-2sigma threshold (what most production dashboards use) serves as baseline. All methods share a single entry point (detect\_anomalies()) and require only a baseline array and specification limits. Parameters use textbook defaults throughout; no tuning is performed between domains.

Experiments

Synthetic data. Three domains with fixed seeds: manufacturing (steel thickness, mu=25.000 mm, sigma=0.010 mm, seed=42), healthcare (ED wait times, mu=28 min, sigma=8 min, seed=123), and AI agent operations (task completion latency, mu=2.5 s, sigma=0.6 s, seed=777). Each domain has a 200-point baseline, 500 monitoring points, and five injected anomaly types.

The AI agent domain models an LLM-based agent processing tasks, with anomalies representing real operational failures: (1) model degradation/API throttling (+1.8sigma mean shift), (2) token explosion/context overflow (2.5x variance), (3) memory leak/state accumulation (linear drift), (4) prompt injection/error spike (+5sigma), and (5) degraded downstream service (+1.4sigma sustained elevation).

SECOM. UCI SECOM dataset(SECOM, 2008): 1,567 semiconductor units, 591 sensors, 6.6% failure rate. Top 10 sensors by variance; first 30% baseline; NaN forward/backward filled.

NAB. Five realKnownCause series from(Ahmad et al., 2017): building HVAC temperature, CPU utilization, EC2 latency, machine temperature, NYC taxi demand. Spans four domains with seven anomaly types. First 15% as baseline. The spc\_engine.py API is applied without modification.

Results

Three-Domain Synthetic Comparison

Cross-domain method comparison (synthetic data, three domains).

Western Electric achieves the highest cross-domain F1 (0.659), a 38% improvement over naive alerting, with average false alarm rate of 3.0% versus 4.7% for naive. The AI agent domain (WE F1!=!0.735) is competitive with healthcare (0.723) and substantially better than manufacturing (0.519), reflecting the cleaner signal structure of agent latency data.

Cross-Domain Consistency (Zero Reconfiguration)

The central empirical finding: the same anomaly-type detection patterns hold across all three domains with identical SPC parameters.

Western Electric region detection rates by anomaly type across domains. Same engine, same parameters, zero tuning. +Fifth anomaly type varies by domain (oscillation/bimodal/sustained).

Three observations. First, the ranking of anomaly difficulty is domain-invariant: spikes are easiest (90--100%), mean shifts next (70--92%), drifts moderate (52--61%), and variance-class changes hardest (10--18%). This reflects the mathematical structure of SPC rules, not domain-specific tuning. Second, detection delays are consistently short for location anomalies (0--2 points for spikes, 2--15 for shifts and drifts) across all domains, despite measurement scales spanning four orders of magnitude (10 mum in manufacturing, 8 min in healthcare, 0.6 s in agent operations). Third, the SPC engine requires only two domain-specific inputs: a baseline array and specification limits. Yet it produces calibrated detection across all three domains. This validates the domain-agnostic claim: the same code that monitors factory sensors monitors AI agents. In the agent domain specifically, catastrophic events (prompt injection) are detected with zero delay, while insidious degradation (memory leaks, tool slowdowns) is detected within 2--15 points. These are operationally useful response times for automated remediation.

SECOM and NAB Validation

On SECOM (real manufacturing), Western Electric achieves F1!=!0.070, +40% over naive, with zero code modification. Low absolute F1 reflects label mismatch (per-unit, not per-sensor); relative advantage is consistent.

On real NAB data (five series, bundled for offline execution), the domain-sensitivity finding is stark. On stationary signals with clear changepoints, Shewhart performs well: ambient temperature system failure (F1!=!0.614), machine temperature (F1!=!0.293). On non-stationary or periodic signals, SPC struggles: NYC taxi demand (WE F1!=!0.074, below naive's 0.175), where zone rules fire on natural periodicity (61% false alarm rate). Shewhart's simplicity is advantageous on real-world data. By flagging only extreme points, it avoids pattern-matching artifacts that plague Western Electric on non-stationary series. Shewhart achieves the best aggregate NAB F1 (0.193, +44% over naive). NAB evaluation uses custom point-level scoring with adaptive windows, not the official NAB scoring profile. Results should be compared within this evaluation framework, not against published NAB leaderboard scores. This reinforces the application rationale: SPC is best suited to stationary operational metrics. This is precisely the signal structure of AI agent latency, and precisely where Western Electric's pattern rules add value rather than noise.

Discussion

SPC vs.\ ML-based monitoring. SPC requires no training data, no labeled anomalies, no API keys, no GPUs. Every alert is explainable (e.g., ``8 consecutive points above center line'') with known false alarm probability. The tradeoff: SPC misses semantic anomalies (e.g., factually incorrect but syntactically valid output). For operational metrics (latency, error rates, throughput), SPC is sufficient and far simpler.

Agent-native monitoring. Because SPC executes in sub-millisecond time per call, it can run inside the agent after each task, flagging degradation before any external system would notice. This is impractical with cloud-based tools that require network calls and authentication. As a proof of concept, SPC-Agent monitors its own execution latency during the pipeline run: 300 iterations of the SPC engine are timed and analyzed by the same engine, producing real (non-synthetic) operational telemetry. The results validate method selection empirically: Western Electric fires over 100 alerts on naturally variable execution latency while Shewhart fires fewer than 5, mirroring the NAB periodic-signal finding. Agent latency monitoring should therefore default to Shewhart for raw metrics, reserving Western Electric for aggregated or preprocessed signals where stationarity holds.

Method selection. We recommend Western Electric as default (highest F1) with CUSUM as supplement for early drift detection (2--15 points earlier). CUSUM's textbook parameters (k=0.5, h=5.0) are tuned for 1\sigma shifts. In a mixed-magnitude anomaly portfolio that includes 4--5\sigma$ spikes, CUSUM's accumulation mechanism underperforms Western Electric's direct pattern rules.

Limitations. Synthetic data uses Gaussian noise (favorable). SECOM labels are per-unit, not per-sensor. SPC assumes stationarity; periodic workloads need detrending. Variance-only anomalies (context overflow) achieve only 10--18% WE detection; MR chart provides dedicated but not dominant coverage. Synthetic anomaly types were chosen to match known SPC detection signatures; cross-domain consistency partly reflects this alignment rather than independent empirical validation.

Reproducibility. All results are deterministic (seeds: 42, 123, 777); verify.py confirms bit-identical reproduction via SHA256 and results.json provides a machine-parseable artifact.

Conclusion

SPC-Agent shows that classical process control, packaged as an executable agent skill, provides effective operational monitoring across manufacturing, healthcare, and AI agent workflows with a single engine, zero reconfiguration, and a single dependency. The cross-domain consistency of detection patterns (Table ) is the key result: anomaly difficulty rankings are invariant to domain, suggesting that SPC's mathematical structure drives detection performance.

Two implications follow. First, any AI agent producing sequential operational metrics can monitor itself using SPC with no external infrastructure. Second, the honest boundary at variance-only anomalies (10--18% detection) and periodic signals defines exactly where SPC should be complemented by other methods, rather than replaced by them.

Next steps are validation on real agent telemetry and streaming SPC for real-time in-agent monitoring. The methods are 70 years old. The application is not.

References

• Ahmad, S., Lavin, A., Purdy, S., & Agha, Z. (2017). Unsupervised real-time anomaly detection for streaming data. Neurocomputing, 262:134--147.
• Qu, Y., Huang, K., Yin, M., et al. (2025). CRISPR-GPT for agentic automation of gene-editing experiments. Nature Biomedical Engineering, 10:245--258.
• Montgomery, D. C. (2012). Introduction to Statistical Quality Control. John Wiley & Sons, 7th edition.
• Page, E. S. (1954). Continuous inspection schemes. Biometrika, 41(1/2):100--115.
• Roberts, S. W. (1959). Control chart tests based on geometric moving averages. Technometrics, 1(3):239--250.
• SECOM (2008). SECOM dataset. UCI Machine Learning Repository. https://doi.org/10.24432/C5RG6N.
• Shewhart, W. A. (1931). Economic Control of Quality of Manufactured Product. D. Van Nostrand Company.
• Western Electric (1956). Statistical Quality Control Handbook. Western Electric Company.