This work presents Aries, a full-stack experimentation framework that separates task semantics from execution configurations, reconstructs cross-component agent trajectories with correlated system telemetry, and exposes stateful tool execution through a consistent interface across heterogeneous sandbox substrates.
Abstract
Autonomous agents challenge conventional LLM serving by coupling repeated inference with persistent context and sandboxed tool execution. We present Aries, a full-stack experimentation framework that separates task semantics from execution configurations, reconstructs cross-component agent trajectories with correlated system telemetry, and exposes stateful tool execution through a consistent interface across heterogeneous sandbox substrates. We use Aries to conduct reproducible experiments on open agent harnesses and benchmarks. We complement these experiments with production traces from a commercial platform, grounding low-level systems research in observed production behavior. Our results show that (1) token-centric metrics miss non-inference bottlenecks, (2) retaining additional context yields diminishing accuracy benefits while reducing serving capacity, and (3) tool sandboxes alternate between long idle periods and short resource bursts, while current snapshot-based state management makes aggressive suspension costly. A complementary security analysis further highlights the need to reduce the sandbox attack surface. We then discuss the vision for agent-native serving systems designed around trajectory-level metrics, adaptive context management, elastic sandbox resource management, and sandboxes with minimized attack surface.
Experimental results show that the proposed self-calibrating agentic framework successfully profiles the zero-knowledge workloads, achieving a higher accuracy than baseline LLM agents and establishing a robust foundation for deploying autonomous AI in decentralized infrastructures.
Fin Gentzen, Marla Grunewald, Iulisloi Zacarias et al.· 0 citations
This paper argues that the community needs a simulation-driven methodology to systematically discover and prevent agent-induced failures across the data sys-tem stack before facing production incidents, and demonstrates this approach on Execution Control Systems (ECS).
Agentic applications are shifting AI serving from isolated model inference to long-running workloads in which LLMs coordinate tools, environments, and persistent state. However, the system behavior of these workloads---where latency, cost, and bottlenecks arise---remains poorly characterized, leaving serving systems to rely on assumptions built for conventional inference. We present AgentSysBench, a benchmark suite and measurement toolkit with ten representative agentic applications and unified systems-level instrumentation. Across controlled deployments and production traces, we identify six properties that distinguish agentic workloads from conventional LLM serving: (1) execution is heavyweight and stateful, with non-LLM components dominating latency in 5 of 10 applications and sandbox working-set memory peaking at 28 GB per session; (2) applications compose components with heterogeneous resource affinity---GPU-bound inference, memory-bound retrieval, CPU-bound sandboxes---whose task latencies diverge by up to 32x; (3) bottlenecks shift across requests, models, and deployments; (4) production sessions hold state idle for minutes to hours between active steps; (5) a control-plane tax---auxiliary LLM calls and context overhead from tool schemas and observations---crowds out productive compute and context; and (6) production traces from three applications reveal heavy cross-request redundancy in search queries and web fetches, exposing a large caching opportunity. Four design explorations demonstrate that these findings are actionable: task-aware serving reduces latency by 29--40%, communication-aware placement by up to 4.5x, state offloading reduces memory usage by 4.6x, and tool-result caching removes 35.2% of redundant search calls and saves 19.3% of aggregate search latency.
Chaokun Chang, Yukun Zhou, Kaihua Fu et al.· 1 citation
Simthesizer introduces a composable simulator infrastructure that uniformly expresses the complete serving workflow, including the control decisions that coordinate it, and realizes it as a unified dynamic graph in Simthesizer simulator.
Wonung Kim, Hyunmin Choi, Minsu Kim et al.· 0 citations
Summary Agentic AI platforms enable the engineering of autonomous workflows but are not designed for experimentation and hypothesis testing. ASAREE (Analytical Sandbox for Agentic AI Research, Engineering, and Experimentation), is an open-source platform to address this gap. ASAREE creates agents, connects to MCP servers and tools, and designs factorial experiments through a visual interface or Python SDK. It records a full provenance trace for every run and routes all model calls through a provider-agnostic bridge that supports local deployments, ensuring data privacy. As a use-case, we use ASAREE to evaluate key design choices in a mutli-agent machine learning pipeline. Across a 2 × 2 × 2 factorial design, more advanced models, greater reasoning effort, and critic agent use significantly increased compute time, token use, cost, and feature count without improving predictive performance. The lowest-cost baseline, Claude Sonnet 5 with medium effort and no critic, achieved the highest mean PR AUC while Claude Opus 5 with extra high effort and a critic agent cost 15.5× more (USD) and ran 13.1× longer while performing worse on average. These findings highlight ASAREE as a robust framework for evaluating agentic system performance and resource efficiency. Availability and implementation ASAREE is available on GitHub at: https://github.com/EpistasisLab/ASAREE. Contact jason.moore@csmc.edu Supplementary Information Supplementary information is available at https://github.com/EpistasisLab/ASAREE/tree/main/publications/bioinformatics
Jay Moran, P. Freda, Attri Ghosh et al.· bioRxiv· 0 citations
The Efficiency-Adjusted Success Rate (EASR) is reported - success weighted by resource consumption relative to declared budgets - revealing that high accuracy at unbounded cost is not production-viable.
Meher Bhaskar Madiraju, Meher Sai Preetam Madiraju· 0 citations