ReliabilityBench: Evaluating LLM Agent Reliability Under Production-Like Stress Conditions

TL;DR

ReliabilityBench addresses the gap between benchmark results and production reliability for tool-using LLM agents by introducing a three-dimensional reliability surface $R(k, \varepsilon, \lambda)$ that captures consistency, robustness, and fault tolerance. The framework leverages Action Metamorphic Relations to define end-state equivalence under perturbations and applies chaos-engineering-style fault injection across four realistic domains, enabling systematic stress testing. Empirical results show perturbations degrade reliability, rate limiting being particularly damaging, and simpler architectures like ReAct offering greater resilience under combined stress, with Gemini 2.0 Flash delivering GPT-4o-comparable reliability at a fraction of the cost. The work demonstrates a practical pathway to production-ready evaluation, emphasizing stress testing, end-state verification, and fault-tolerant design as essential components of LLM agent deployment. The approach provides a rigorous, scalable method to quantify and compare reliability across models, architectures, and domains, informing deployment decisions and reliability engineering in real-world systems.

Abstract

Existing benchmarks for tool-using LLM agents primarily report single-run success rates and miss reliability properties required in production. We introduce \textbf{ReliabilityBench}, a benchmark for evaluating agent reliability across three dimensions: (i) consistency under repeated execution using $\mathrm{pass}^k$, (ii) robustness to semantically equivalent task perturbations at intensity $ε$, and (iii) fault tolerance under controlled tool/API failures at intensity $λ$. ReliabilityBench contributes a unified reliability surface $R(k,ε,λ)$, \textit{action metamorphic relations} that define correctness via end-state equivalence rather than text similarity, and a chaos-engineering-style fault injection framework (timeouts, rate limits, partial responses, schema drift). We evaluate two models (Gemini 2.0 Flash, GPT-4o) and two agent architectures (ReAct, Reflexion) across four domains (scheduling, travel, customer support, e-commerce) over 1,280 episodes. Perturbations alone reduce success from 96.9% at $ε=0$ to 88.1% at $ε=0.2$. Rate limiting is the most damaging fault in ablations. ReAct is more robust than Reflexion under combined stress, and Gemini 2.0 Flash achieves comparable reliability to GPT-4o at much lower cost. ReliabilityBench provides a systematic framework for assessing production readiness of LLM agents.

Figures (5)

• Figure 1: Action Metamorphic Relation: Task description perturbation $\phi$ should preserve goal satisfaction. The agent may take different actions but must achieve equivalent end states.
• Figure 2: Reliability degradation under fault injection at measured $\lambda$ levels. ReAct shows 7.5% degradation from $\lambda=0$ to $\lambda=0.2$, while Reflexion shows 10.0% degradation, indicating that self-reflection mechanisms may amplify rather than mitigate fault impacts.
• Figure 3: Measured Reliability Surface $R(k{=}2, \varepsilon, \lambda)$ for Gemini 2.0 Flash. Each point shows the pass$^2$ rate at the measured $(\varepsilon, \lambda)$ grid point. Baseline ($\varepsilon{=}0, \lambda{=}0$) achieves 96.88%, degrading to 84.0% under combined perturbation and fault stress.
• Figure 4: Reliability degradation under increasing perturbation levels. Both models show $\sim$8-9% drop from baseline ($\varepsilon=0$) to medium perturbation ($\varepsilon=0.2$), with the steepest drop occurring at $\varepsilon=0.1$.
• Figure 5: Fault type ablation results. Rate limiting causes the largest degradation (93.75%), while transient timeouts are best handled (98.75%).

Theorems & Definitions (9)

• Definition 1: Agentic Task
• Definition 2: Agent Execution
• Definition 3: pass$^k$
• Definition 4: Perturbation Level $\varepsilon$
• Definition 5: Fault Intensity $\lambda$
• Definition 6: Reliability Surface
• Definition 7: Action Metamorphic Relation
• Definition 8: Fault Type
• Definition 9: Fault Profile