Characterizing Faults in Agentic AI: A Taxonomy of Types, Symptoms, and Root Causes
Shah et al. (arXiv:2603.06847) · v2026-03 · first published 2026-03-06
Scope
Grounded-theory taxonomy of faults in agentic AI systems, derived from 13,602 issues and PRs across 40 open-source agentic-AI repositories, with stratified sampling to 385 faults and developer-study validation (n=145). Three parallel taxonomies: fault types (5 dimensions, 13 major categories, 37 fault subcategories); 13 symptom classes; 12 root-cause categories. This is the empirical corpus SAICA-KG uses for its v0.1 back-test.
Why it's ingested
Shah et al. 2026 is the single most comprehensive empirical fault characterization for agentic AI to date, with a 385-fault grounded-theory sample across 40 real repos — exactly the scale SAICA-KG needs to argue its FailureMode partition is honest-MECE. Intended crosswalks: SHAH-MAJ-10 (Dependency Management) -> obsolescence + supply_chain_attack; SHAH-MAJ-03 (Tool Execution/API Usage) -> logic_error + fabrication (API); SHAH-MAJ-08 (Context/State Persistence) -> context_pollution; SHAH-MAJ-09 (Input Interpretation/Logic) -> logic_error; SHAH-RC-04 (Auth/Credential) -> security_vulnerability; SHAH-RC-08 (API Schema Drift) -> obsolescence. Coverage ratio against SAICA-KG's FailureMode enum is the v0.1 honest-MECE evidence.
Categories & crosswalks
Each row lists one external category and any SAICA-KG FailureModes that cross-walk into it.
SHAH-DIM-01Agent Cognition and OrchestrationFailures in cognitive architecture and control flows unique to agentic systems (83 faults in the sample). Covers LLM integration and agent lifecycle/state.
SHAH-DIM-02Tooling, Integration, and ActuationDefects in tool invocation and external-system interaction (66 faults). Covers API usage, external connectivity, resource manipulation, coordination, and observability of tool calls.
SHAH-DIM-03Perception, Context, and MemoryFaults in information ingestion, context preservation, and input interpretation (72 faults). Covers persistence and input-interpretation logic.
SHAH-DIM-04Runtime and Environment GroundingExecution environment, dependency stack, and platform-constraint mismatches (87 faults). The single largest fault dimension by count, dominated by dependency management.
SHAH-DIM-05System Reliability and ObservabilityRobustness deficiencies and diagnostic challenges in autonomous systems (67 faults). Covers error recovery, UI/visualization, and documentation.
SHAH-MAJ-01LLM Integration FaultsFaults in how the agent integrates with its LLM (45 faults).
SAICA-KG crosswalks:SHAH-MAJ-02Agent Lifecycle and StateFaults in agent startup, execution, termination, and state tracking (38 faults).
SAICA-KG crosswalks:SHAH-MAJ-03Tool Execution and API UsageIncorrect or malformed tool/API invocations (17 faults).
SAICA-KG crosswalks:SHAH-MAJ-04External Connectivity and AccessConnection setup and authentication failures (19 faults).
SAICA-KG crosswalks:SHAH-MAJ-05Resource ManipulationResource-handling and database-misconfiguration faults (16 faults).
SAICA-KG crosswalks:SHAH-MAJ-06System CoordinationSynchronization errors across components or agents (6 faults).
SHAH-MAJ-07System ObservabilityLogging or telemetry defects (8 faults).
SHAH-MAJ-08Context and State PersistenceMemory-persistence and state-load/save failures (12 faults).
SAICA-KG crosswalks:SHAH-MAJ-09Input Interpretation and LogicType handling, validation, encoding, and logic/constraint violations (60 faults).
SAICA-KG crosswalks:SHAH-MAJ-10Dependency ManagementDependency specification, import resolution, installation, and package resolver/manager failures (67 faults). The single largest major category.
SAICA-KG crosswalks:SHAH-MAJ-11Platform CompatibilityEnvironment, platform, API, and integration compatibility mismatches (20 faults).
SAICA-KG crosswalks:SHAH-MAJ-12Robustness and Error RecoveryException-handling and implementation defects (49 faults).
SAICA-KG crosswalks:SHAH-MAJ-13UI, Visualization, and Documentation DefectsUI defects (13 faults) and documentation defects (5 faults).
SHAH-RC-01Dependency and Integration FailuresContract violations between probabilistic artifacts and deterministic interfaces (19.5%).
SAICA-KG crosswalks:SHAH-RC-02Data and Type Handling FailuresType mismatches and validation gaps (17.6%).
SAICA-KG crosswalks:SHAH-RC-03Incorrect Token ManagementTracking and counting errors affecting context windows.
SAICA-KG crosswalks:SHAH-RC-04Authentication and Credential IssuesFragile token refresh and credential-expiration handling.
SAICA-KG crosswalks:SHAH-RC-05Datetime Conversion ErrorsInappropriate temporal-value transformations.
SAICA-KG crosswalks:SHAH-RC-06State Management DefectsInconsistent internal state across interactions.
SAICA-KG crosswalks:SHAH-RC-07Configuration InconsistenciesWrong defaults and misconfiguration.
SAICA-KG crosswalks:SHAH-RC-08API Evolution and Schema DriftIncompatibilities from upstream API changes.
SAICA-KG crosswalks:SHAH-RC-09Environment-Level IssuesPlatform-specific constraints and virtual-environment problems.
SAICA-KG crosswalks:SHAH-RC-10Documentation GapsMissing or incorrect configuration guidance.
SHAH-RC-11Exception SuppressionSwallowed errors reducing observability.
SAICA-KG crosswalks:SHAH-RC-12Ecosystem FragilityRapid upstream evolution creating tight-coupling vulnerabilities.
SAICA-KG crosswalks:
Notes
Top-level structure (5 dimensions, 13 major categories, 12 root causes) is canonical per the arXiv v1 HTML; the 37 fault subcategories are enumerated in the paper and will be expanded into full category rows as a v0.2 task. Symptom-class and full subcategory labels are pending full-text verification. This taxonomy is the empirical back-test corpus: SAICA-KG's FailureMode axis will be evaluated against whether SHAH-* faults map cleanly into its eight-class FailureMode enum.