How Fast Do Agents Rot? An Empirical Study of Long-Horizon Degradation in LLM Agents for Production Decision-Making

Production deployments of large language model (LLM) agents remain unreliable on long, multi-step workflows even as benchmark success rates climb steadily. We argue this gap is largely an artifact of task horizon: benchmarks are dominated by short-to-medium horizons where success remains high, while production workloads demand an order of magnitude more dependent steps. We measure the effect directly, characterizing the shape of agent degradation and disentangling its cause across a large controlled study spanning nine models, six open models from 1.2B to 671B parameters, and three deployed proprietary systems; four task families, including a genuinely agentic tool-use loop; five horizons; and three context regimes. Task success follows a geometric law governed by a single per-step reliability parameter, which rises with model scale but saturates well below 1 even for the strongest models, guaranteeing eventual collapse at sufficiently long horizons. The effect is sharpest on the agentic task, where every model tested, including widely deployed systems, falls from near-perfect success to near zero within sixteen steps of (n=10,664 analyzed trajectories. Degradation is driven by step count rather than context length: bounding the context window steepens decay rather than easing it (logit slope -0.69 vs. -0.44), p=3x10-6), contradicting a lost-in-the-middle explanation and warning against a common production shortcut. Projecting measured reliability onto representative benchmark horizons quantifies a substantial gap between benchmark and production conditions, from 0.42 at GAIA-length horizons to 0.24 at hundred-step production horizons. For teams responsible for agent orchestration and reliability at scale, these results argue for horizon-aware evaluation and reliability budgeting in place of aggregate pass-rate metrics. Code, prompts, seeds, and raw trajectories are released.

Beyond Capability Benchmarks: Learning Operational Fingerprints of LLM Cloud Services from Production Incident Metadata

Managed LLM services are now part of real production systems, but model selection and service planning still rely heavily on capability benchmarks that reveal little about operational behavior after deployment. We present Operational Embedding (OpEmbed), a framework for learning compact operational fingerprints of LLM cloud services from structured, privacy-preserving support-case metadata, without using case text. OpEmbed aggregates model--time windows into an eight-channel operational signature and learns a low-dimensional representation via temporal contrastive learning, cross-view reconstruction, and generational-ordinality regularization. Evaluated on more than 33,000 production support cases spanning seven LLM families over 26 months at Google Cloud, OpEmbed recovers interpretable family- and version-level structure, improves leave-one-model-out operational forecasting over non-learned baselines, remains useful under limited early-window data, and supports cross-model fault-type transfer. We report the practical lessons learned from building and evaluating this tool for model onboarding, support readiness assessment, and operational monitoring.

TurboBias 2.0: Streaming Context-Biasing for Production-Efficient ASR Systems

Contextualization is essential for production automatic speech recognition (ASR) systems, where user-provided phrases must be recognized accurately under strict latency constraints. Although many context-biasing methods improve recognition accuracy, they often do not address the practical requirements of modern production ASR systems: streaming inference, efficient batched decoding, user-specific context lists, and low runtime overhead. We propose TurboBias 2.0, a production-oriented framework for efficient phrase boosting in Transducer-based ASR systems. The framework extends GPU-accelerated TurboBias with a case-insensitive boosting graph and per-stream batched decoding, allowing each utterance in a batch to use an independent context-biasing configuration. This enables personalized context biasing for multiple simultaneous users without sharing or mixing their context lists. The proposed framework supports both offline and streaming inference and can be used with greedy and beam-search decoding. Experiments show that TurboBias 2.0 improves contextual phrase recognition while preserving low latency and high throughput.

Governed Shared Memory for Multi-Agent LLM Systems

Multi-agent LLM environments require robust mechanisms for shared knowledge management. This paper formalizes the fleet-memory problem and identifies four foundational failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. To address these, we define explicit systems-level primitives: scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation. These primitives are implemented in MemClaw, a production multi-tenant memory service, and evaluated via ArgusFleet, a reproducible harness testing four governance dimensions. Rather than a baseline comparison, this study measures a live production service, emphasizing real-world architectural insights and negative results. Key Evaluation Results Provenance: Successfully reconstructed 100% of depth-four derivation chains with correct writer identity at sub-second per-hop latency. Propagation: Demonstrated high intra-fleet visibility with zero cross-fleet leakage. Under strong write mode, write-to-visible latency was optimized to a single search round-trip. Production Architectural Issues Discovered Asymmetric Scope Enforcement: Tenant isolation held, but sub-tenant scope was initially bypassed on direct GET-by-id requests for agent-scoped credentials (disclosed and remediated during the study). Pipeline Ordering Conflict: While contradiction supersession works for admitted writes, a synchronous near-duplicate gate can prematurely reject contradictory writes before the asynchronous contradiction detector can evaluate them. Conclusion: Long-context retrieval alone is insufficient for production multi-agent memory. Governed shared memory demands explicit systems-level abstractions, and live evaluation is vital to expose enforcement and pipeline-ordering failures missed by design-only treatments.

When Do LLMs Replace Fine-Tuned NLU? A Decision Framework for Intent Detection in Production Conversational Systems

A common claim is that zero-shot large language models (LLMs) can replace fine-tuned NLU classifiers for intent detection. We test this claim head-to-head and find that the honest answer is: it depends on the intent space. On full ATIS and CLINC150 we compare a fine-tuned RoBERTa, a TF-IDF+logistic-regression baseline, sentence-embedding kNN, and Claude Haiku zero-shot, reporting bootstrap 95% confidence intervals and paired significance tests. When abundant in-domain labels exist, fine-tuned RoBERTa is as good or better and three orders of magnitude cheaper and faster: on ATIS it beats Claude zero-shot by 11.8 points (95.9 vs. 84.1, p<0.001). On the broad 150-intent CLINC150 schema the two are statistically tied (89.1 vs. 88.5, p=0.24): the LLM matches a fully supervised model with no training data. The LLM's advantages appear in three production-relevant regimes: out-of-scope detection (OOS recall 85.6 vs. 58.1 for RoBERTa); robustness to realistic ASR noise via a controlled text-to-speech to noise to Whisper pipeline (92.5 vs. 80.0 at 0 dB); and dynamic per-deployment schemas, where a classifier trained on one app's intents scores 0% on a new app's intents while the schema-prompted LLM serves both at ~94% with zero retraining. We distill these findings into a decision framework for practitioners.

When Errors Become Narratives: A Longitudinal Taxonomy of Silent Failures in a Production LLM Agent Runtime

LLM agent systems increasingly run as long-lived autonomous runtimes: scheduling jobs, calling tools, maintaining memory, and pushing results to humans. We present a longitudinal study of silent failures in one such system: a personal-assistant agent runtime in continuous production since March 2026, with roughly 40 scheduled jobs, 8 LLM providers, a tool-governance proxy, and a knowledge-base memory plane, defended by 4,286 unit tests and 827 governance checks. Over eight weeks we documented 22 incidents with full root-cause postmortems, in which one meta-pattern -- a failure whose error signal never reaches a human in actionable form -- manifested at least 28 times. We derive a five-class, mechanism-oriented taxonomy: (A) environment and platform quirks, (B) design-assumption mismatches, (C) error swallowing and dilution, (D) chained hallucination and fabrication, (E) operational omission and forensic blind spots. Class D is unique to LLM systems and the most dangerous: the system does not merely fail to report an error -- the LLM transforms it into fluent, plausible narrative delivered to the user. We term this fail-plausible: gray failure's differential observability escalated -- the observer is not just blind, it is convincingly lied to by the failure itself. Three findings: about 70% of silent failures were caught by human user-view observation, not tests or audits; a retrospective audit of 15 incidents found 0% ex-ante prevention but 87% regression blocking -- audits are regression engines, not prediction engines; incident latency (13 hours to 60 days) tracks failure mechanism, not code complexity -- the longest-lived failures lived in the seams between components, where no test runs. We describe the resulting defense framework and distill design principles for agent systems whose failures are loud, attributable, and boring. All postmortems and artifacts are public.

External Experience Serving in Production LLM Systems: A Deployment-Oriented Study of Quality-Cost Trade-offs

Production LLM systems accumulate reusable operational experience, but the practical deployment issue is not merely whether such experience can help. It is how different serving strategies trade off quality against online cost under realistic constraints. Injecting external experience can improve task quality, yet it also increases prompt burden, latency, and serving pressure. We study \textit{external experience serving} as a deployment-oriented quality-cost trade-off problem. We evaluate this question in a real production moderation setting, with tool-use and GPQA as supporting contrast tasks that expose different output-cost regimes. We compare no-experience baselines, random experience controls, global prompt injection, and retrieval-based selective injection, and analyze both task quality and serving cost. The results show that, once experience becomes case-dependent, selective retrieval provides a stronger operating point than unconditional global injection. They further show that retrieval quality matters more than simply increasing Top-$K$, and that the same serving policy can exhibit substantially different cost-benefit profiles across short-output and decode-heavy regimes. These findings suggest that external experience is best treated as a selective, cost-aware serving decision rather than as a universal add-on. Overall, in the settings studied here, external experience pays off only when both the serving interface and the task-specific cost structure make its quality gains worth the online cost.

From Capability Models to Automated Planning: An AAS-Native Approach for Automatic PDDL Generation

Engineers designing production systems need to verify that a given layout supports all required production sequences. Automated planning techniques can answer such questions, but formulating the required planning problems in the Planning Domain Definition Language (PDDL) demands specialized expertise that production engineers typically lack. Asset Administration Shells (AAS) have emerged as the standardized Digital Twin for industrial assets in Industry 4.0. We show that AAS capability models, structured using four established Industry 4.0 standards (VDI 3682 for process descriptions, IEC 61360-1 for semantic property qualification, IDTA 02011 for type hierarchies, and IDTA 02016 for instance descriptions), contain sufficient information to generate complete PDDL problems automatically. Unlike prior work that introduced PDDL-specific submodels, our approach derives all planning elements from domain-level descriptions of resource functions, so-called capabilities, allowing engineers to model capabilities without any exposure to PDDL syntax or planning concepts. Our extraction algorithm transforms distributed Multi-AAS architectures into complete PDDL planning problems. We validate the approach on AAS models of a laboratory production system, comparing four layout variants using optimal planning to demonstrate how engineers can systematically explore design trade-offs by modifying the AAS model and regenerating the planning domain