Agentic Error Propagation in AI Systems

• Agentic error propagation is the cascading chain of errors in multi-step AI systems where minor inaccuracies can lead to significant, irreversible system failures.
• Structural models including directed dependency DAGs, the spiral of hallucination, and multi-agent cascade dynamics provide insights into how errors propagate across interconnected workflows.
• Mitigation strategies such as uncertainty decomposition, dual-process control, and governance instrumentation are crucial for preventing error amplification and maintaining high system integrity.

Agentic error propagation is the process by which errors—whether in beliefs, reasoning, decisions, or outputs—originate and cascade through sequential, multi-step, or multi-agent AI systems, amplifying technical failures into broader, often irrecoverable, harms. In agentic systems, local inaccuracies can become entrenched in internal state or workflow context, contaminating downstream planning, fostering false consensus, triggering distributed or structural failures, and undermining both safety and accountability. This phenomenon is now a central challenge for the reliability, calibration, and governance of high-autonomy AI deployments across domains such as finance, code generation, workflow automation, collaborative research, and real-time control.

1. Mathematical Foundations and Formal Risk Decompositions

A central analytic framework for agentic error propagation is the Bayesian decomposition of automation risk, as developed by Srivastava & Sah. The expected loss per decision, reflecting the compounding of model and execution errors, is given by

$$\mathbb{E}[\mathrm{Loss}] = P(F) \times P(H \mid F, A) \times \mathbb{E}[S \mid H]$$

where:

• $P(F)$ is the marginal probability of system (model) failure ("technical risk"),
• $P(H\mid F, A)$ is the conditional probability that a failure propagates into harm given automation level $A$ ("deployment risk"),
• $\mathbb{E}[S\mid H]$ is expected harm severity ("consequence risk") (Srivastava et al., 22 Feb 2026).

The harm propagation equivalence theorem states that, under natural conditions (harm only occurs via executed failures), the harm-propagation term reduces to the probability that a failed output is actually executed, which connects inherently unobservable risk to observable system controls such as override rates and kill-switch latencies.

Sensitivity to automation is quantified by elasticities, e.g.,

$$\varepsilon_{A} = \frac{A}{P(H\mid F,A)} \frac{\partial P(H\mid F,A)}{\partial A}$$

so that irreversible agentic control (high $A$) sharply magnifies risk amplification. These analyses enable Pareto-efficient risk–cost tradeoffs and optimal allocation of validation or oversight resources, as exemplified by the empirically calibrated Knight Capital case (\$440M loss, nearly unitary propagation probability).

2. Structural Models of Error Propagation in Agentic Workflows

Agentic systems frequently manifest as orchestrated workflows or distributed multi-agent ecologies, with cascading dependencies and non-trivial failure chains. Structural analysis identifies several core forms of propagation:

• Directed Dependency DAGs: Agent executions are best modeled as evaluation DAGs $\mathcal{G} = (V, E, \tau, \mathcal{M})$, where edges define the propagation pathways for failures (Guo et al., 26 Apr 2026). Root-cause errors in upstream nodes are automatically attributed to downstream failures that consume their outputs.
• Spiral of Hallucination: In long-horizon reasoning, the trajectory’s validity probability evolves as

$P(V_{t} = 1) \approx \prod_{i=1}^{t} c_{i}$

where $c_i$ is local confidence. A single catastrophic error at step $P(F)$0 drives the entire chain of validity to zero for $P(F)$1, rendering the error irreversible unless proactively corrected (Zhang et al., 22 Jan 2026).

• Multi-Agent Cascade Dynamics: Propagation in collaborative multi-agent systems is governed by mean-field update equations over the adoption of atomic falsehoods. For agent $P(F)$2 at round $P(F)$3:

$P(F)$4

where $P(F)$5 is the per-hop propagation probability and $P(F)$6 the self-correction rate. Supercritical regimes ($P(F)$7) exhibit rapid amplification and false consensus (Xie et al., 4 Mar 2026).

3. Empirical Taxonomies and Propagation Pathways

Large-scale empirical studies reveal that propagation pathways are not random but follow reproducible patterns. Key findings include:

• Fault Associations: Statistical mining of 385 real-world faults in agentic AI systems demonstrates that specific error types and symptoms predictably trigger distinct downstream faults, e.g., token-tracking errors → authentication failures (lift=181.5), naive datetime encoding → scheduling anomalies (lift=121.0), and memory persistence bugs → context loss (Shah et al., 6 Mar 2026).
• Hierarchical Causality: Most error cascades observed in practice emerge not from independent random faults but from systematic mismatches at module interfaces—such as probabilistic artifacts (LLM-generated code) clashing with deterministic parsing or tool requirements.
• Propagation Examples: Case studies detail how a single misalignment in a token-counting logic can cause authentication failures across downstream workflow stages and why naive datetime usage propagates to system-wide scheduling errors.

Upstream Fault	Propagated Symptom	Pathway Confidence	Pathway Lift
Token counting error	API auth failure	1.00	181.5
Naive datetime encoding	Scheduling anomaly	1.00	121.0
Memory write bug	Context loss	1.00	30.25

Such empirical mapping supports the development of structured debugging and automated root-cause attribution frameworks.

4. Quantifying, Diagnosing, and Mitigating Propagation

Methodological advances address both quantification and mitigation of agentic error propagation:

• Uncertainty Decomposition: Internal (intrinsic) and external (extrinsic) uncertainties can be separated using information-theoretic tools. UProp leverages the chain rule of entropy to estimate extrinsic uncertainty as the mutual information inherited from prior decisions, allowing selective deferral to a human or risk-aware decision thresholds (Duan et al., 20 Jun 2025).
• Concentration Inequalities: For tool-using agents, martingale analysis establishes that cumulative semantic distortion grows at most linearly in step count, with high-probability deviations bounded by $P(F)$8 provided the branching factor and context-decay satisfy stability criteria. Periodic re-grounding (approximately every 9 steps for typical context-decay) suffices to control error (Fan et al., 10 Feb 2026).
• Dual-Process Control: Agentic Uncertainty Quantification frameworks (e.g., AUQ) combine an uncertainty-aware memory for forward propagation with an explicit reflection mechanism for targeted correction. Dynamic thresholds prevent over-reflection and enable a balance between efficiency and error arrest (Zhang et al., 22 Jan 2026).
• Governance Instrumentation: The "cascade of uncertainty" in agentic governance is addressed through distributed trace protocols, evidence aggregation, and delegated responsibility records to restore accountability after evidence fragmentation and decision diffusion (Solozobov, 21 Apr 2026).

5. Architectural and Systemic Countermeasures

Mitigation strategies span multiple levels—architecture, workflow, and communication:

• Genealogy-Graph Governance: Message-layer plugins construct a lineage graph for atomic claims, triaging each for verification, blocking, or approval. This message-level gating reduces the per-hop propagation parameter and empirically raises benign infection control from 0.32 to 0.89 in deployment (Xie et al., 4 Mar 2026).
• Exploration-Stage Redundancy: The ExComm protocol introduces inter-agent belief auditing for factual conflicts, soft belief updates for correction without over-synchronization, and explicit trajectory diversification to prevent collapse of reasoning diversity, yielding substantial accuracy and recovery gains (Song et al., 21 May 2026).
• Progressive Error Feedback: Iterative, agentic error feedback with escalating complexity, as in PEFA-AI for code generation, accelerates convergence and suppresses error entrenchment while outpacing non-agentic baselines in both pass rates and token efficiency (Narayanan et al., 6 Nov 2025).
• Directed Acyclic Graph Evaluations: Step-level DAG evaluation (AgentEval) enables precise failure detection and root-cause attribution, restoring the ability to distinguish new faults from propagation and delivering a >2x recall increase over end-to-end or flat stepwise checks (Guo et al., 26 Apr 2026).

6. Open Problems and Future Directions

Despite these advances, significant challenges and research directions remain:

• Long-horizon Consistency: Maintaining anchor state across multi-hop, memory-constrained chains—critical for reasoning and open-domain retrieval—is an unsolved bottleneck, mitigated only partially by "workflow" style explicit memories or partial re-grounding (Liu et al., 11 Jan 2026).
• Agentic Governance Boundaries: Structural breaks such as responsibility ambiguity, evidence fragmentation, and decision diffusion in dynamic agentic systems undermine traditional governance. Extension protocols supply only partial remediation and depend on white-box access and cooperation, leaving boundary conditions (propagating opacity, black-box endpoints) as open limitations (Solozobov, 21 Apr 2026).
• Domain-agnostic Evaluation: There is strong experimental evidence that DAG-based and message-graph metrics generalize across domains, but real-world agentic systems increasingly interleave structured DAGs with unstructured or looped workflows, requiring more general tracking mechanisms (Guo et al., 26 Apr 2026).
• Robust Interpretability: System-level interpretability—tracing propagation and error chains in live deployments—remains an open research agenda. Standardizing provenance, causal dashboards, and regulatory traceability are ongoing priorities to support both detection and prevention of catastrophic propagation (Zhu et al., 23 Jan 2026).

By unifying analytic risk decompositions, structural models, empirical taxonomies, and targeted countermeasures, current research delineates both the scope and mitigation of agentic error propagation, offering principled pathways for advancing reliability in high-autonomy AI systems.