Recoleta
GitHub
Source note

Monitoring Agentic Systems Before They're Reliable

Authors
Marisa Ferrara Boston, Glen Hanson, Effi Georgala, JD Hudgens, Heather Frase
Source
arXiv
Published
Collection
Software Intelligence
Agentic SystemsAgent MonitoringAI OpsFmea TriageReliability Engineering
Open arXivMarkdown source

Summary

This paper claims early production agentic systems need structural monitoring before task-level error detection works. It tests a variance-based monitoring and FMEA triage method on a synthetic audit-agent testbed and finds that monitor scope predicts the kind of failure found.

Problem

  • Early agentic systems often fail because stages, tools, or data paths are poorly integrated, so task-level monitors may miss the real cause.
  • Mean accuracy and threshold alerts can hide rare high-risk failures in regulated workflows such as audit, finance, healthcare, and legal services.
  • Human reviewers need severity routing because raw monitoring findings can create more review work than teams can handle.

Approach

  • The method scores agent behavior across three dimensions: quality, suitability, and efficiency.
  • It uses three monitoring scopes: within-run monitors for variation inside one execution, cross-run monitors for variation across repeated executions, and structural monitors for integration gaps.
  • Variance is a main signal. The paper uses z-scores with thresholds of 2.0 for most evaluators and 3.0 for timing-sensitive evaluators, plus coefficient of variation to characterize monitor classes.
  • Findings are sent into an FMEA-style triage process with four severity levels: L1 catastrophic, L2 critical, L3 marginal, and L4 negligible.
  • The evaluation uses 120 synthetic audit document bundles: 20 clean bundles and 100 error bundles across 5 error subtypes, 4 difficulty levels, and 5 document sets per cell, processed in 220 runs.

Results

  • Within-run monitors found deterministic stage defects with low variation: CV = 0.02. These produced high-volume, low-severity L3 findings.
  • Cross-run monitors found stochastic integration consequences with high variation: CV = 1.25, with 24% of findings classified as L2 critical.
  • The structural monitor found an integration gap with perfect consistency: CV = 0.00 across runs.
  • Injected task-level errors were statistically indistinguishable from clean baselines, including 100 injected errors, supporting the claim that structural defects can mask task-level signals in immature systems.
  • Deterministic triage routed all 10,210 L3 findings to automated monitoring and all 243 L2 findings to human investigation, reported as a 43x reduction in analyst review volume.
  • The abstract also reports that 97% of findings went to automated tracking and about 2% went to human investigation for variable system behavior.

Link