pith. sign in

arxiv: 2605.26297 · v1 · pith:XMXJDBPVnew · submitted 2026-05-25 · 💻 cs.DC

Agentic AI Workload Characteristics

Yichao Yuan , Ankita Nayak , Souvik Kundu , Nishil Talati This is my paper

Pith reviewed 2026-06-29 20:09 UTC · model grok-4.3

classification 💻 cs.DC
keywords agentic AILLM servingcontext cachingKV-cachetool useReAct agentsmulti-turn executionworkload characterization
0
0 comments X

The pith

Agentic AI workloads become decode-dominated with context caching because most input tokens are reused across turns.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper traces ReAct-style agents on five benchmarks using Gemma and Qwen models to map how agentic AI changes from isolated prompt requests to stateful multi-turn executions that invoke the model repeatedly and grow context. With effective context caching the bulk of tokens are reused, so the workload tilts toward decoding and places heavier demands on long-lived KV-cache state. Tool calls also follow a temporal pattern that moves from read and explore early in a run to execute and write later. These traits matter because serving systems built for single-shot prompts will need new mechanisms to handle repeated re-entry and persistent state. The study therefore supplies concrete workload data that any design for agentic serving must accommodate.

Core claim

Agentic AI shifts LLM serving from isolated prompt-generation requests to stateful, multi-turn executions that repeatedly invoke the model, call tools, and grow context over time. With effective context caching, most input tokens are reused across turns, making execution decode-dominated while increasing dependence on long-lived KV-cache state. Tool use has a clear temporal structure, with agents shifting from read/explore behavior early in execution to execute/write behavior later.

What carries the argument

End-to-end tracing of ReAct-style agent executions that records both LLM calls and tool invocations across turns on multiple benchmarks.

If this is right

  • Agentic workloads are not simply long-prompt workloads once caching is applied.
  • Serving systems must jointly manage repeated model re-entry, persistent KV-cache state, and workload-dependent tool behavior.
  • Decode phases dominate execution time, raising the relative cost of KV-cache residency.
  • Tool-use patterns change over the lifetime of an agent run, so resource allocation can be phased accordingly.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Hardware designs that favor high decode throughput over prefill throughput may gain an advantage for agentic traffic.
  • Caching policies could be tuned to the observed read-to-write transition rather than treating all context uniformly.
  • Future benchmarks should include explicit measurement of token reuse rates and tool-phase timing to remain representative.

Load-bearing premise

The five chosen benchmarks and the ReAct-style pattern on Gemma and Qwen models stand in for the wider range of agentic workloads that will appear in production.

What would settle it

A new agentic benchmark suite or different model family that shows low token reuse under caching or lacks the early-read to late-write tool shift would falsify the reported workload characteristics.

Figures

Figures reproduced from arXiv: 2605.26297 by Ankita Nayak, Nishil Talati, Souvik Kundu, Yichao Yuan.

Figure 1
Figure 1. Figure 1: ReAct agent execution and context growth example. Each turn consists of an LLM response, one or more tool [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Tracing infrastructure for characterizing agent [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of agent turns per task. Agent trajectories are highly variable: instant variants often require more [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of accumulated context usage across all turns for each task. Context usage varies substantially by [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Breakdown of generated output tokens into [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Context-usage distributions split by task outcome. Failed agents often carry larger contexts (up to 1.8 [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Breakdown of end-to-end execution time between [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Breakdown of per-turn context into cached input tokens, newly appended input tokens, and output tokens. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Breakdown of LLM execution time into prefill and [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 7
Figure 7. Figure 7: its tools are not merely frequent, but also include [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 10
Figure 10. Figure 10: Breakdown of tool-call types across workloads. Tool usage is highly model and domain-dependent: coding and [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Breakdown of Bash commands issued by agents. Command usage reflects the task domain: database-oriented [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Progression of high-level tool intent over agent execution. Most agents shift from read/explore-heavy behavior [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
read the original abstract

Agentic AI shifts LLM serving from isolated prompt-generation requests to stateful, multi-turn executions that repeatedly invoke the model, call tools, and grow context over time. This paper characterizes ReAct-style agents from both the LLM-serving and tool-execution perspectives using an end-to-end tracing infrastructure across reasoning and non-reasoning Gemma and Qwen configurations on five agentic benchmarks. Our study shows that agentic workloads are not simply long-prompt workloads: with effective context caching, most input tokens are reused across turns, making execution decode-dominated while increasing dependence on long-lived KV-cache state. We also find that tool use has a clear temporal structure, with agents shifting from read/explore behavior early in execution to execute/write behavior later. These results show that efficient agentic serving must jointly manage repeated model re-entry, persistent context state, and workload-dependent tool behavior.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript characterizes ReAct-style agentic AI workloads via end-to-end tracing on five benchmarks using Gemma and Qwen models (reasoning and non-reasoning configurations). It claims that, with effective context caching, most input tokens are reused across turns (making execution decode-dominated and increasing dependence on long-lived KV-cache state) and that tool use exhibits a clear temporal structure, shifting from read/explore behavior early in execution to execute/write behavior later. These observations are used to argue that efficient agentic serving must jointly manage repeated model re-entry, persistent context state, and workload-dependent tool behavior.

Significance. If the traced patterns hold and prove representative, the work would provide actionable guidance for LLM serving systems targeting stateful multi-turn agents, particularly around KV-cache management and phase-aware scheduling. The end-to-end tracing infrastructure is a positive methodological contribution that supports reproducible measurement of these workload characteristics.

major comments (2)
  1. [Abstract] Abstract: the central claims that 'most input tokens are reused across turns, making execution decode-dominated' and that tool use has a 'clear temporal structure' are stated at a high level with no accompanying quantitative data (e.g., reuse ratios, token counts per phase, or statistical summaries) from the traces on the five benchmarks. This absence prevents evaluation of whether the measurements support the stated conclusions.
  2. [Abstract] Abstract (final sentence): the serving implications ('efficient agentic serving must jointly manage repeated model re-entry, persistent context state, and workload-dependent tool behavior') are drawn from ReAct-style traces on five specific benchmarks with Gemma/Qwen models. No analysis or discussion is supplied to establish that these execution graphs, control flows, or context-management strategies are representative of broader agentic workloads (e.g., parallel tool invocation or hierarchical planning), which could materially alter the reported token-reuse and phase-shift statistics.
minor comments (1)
  1. [Abstract] The abstract would be strengthened by briefly noting the number of traces, model scales, or key quantitative highlights to give readers an immediate sense of the data supporting the high-level findings.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight opportunities to strengthen the presentation of our quantitative results and the scope of our claims. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claims that 'most input tokens are reused across turns, making execution decode-dominated' and that tool use has a 'clear temporal structure' are stated at a high level with no accompanying quantitative data (e.g., reuse ratios, token counts per phase, or statistical summaries) from the traces on the five benchmarks. This absence prevents evaluation of whether the measurements support the stated conclusions.

    Authors: We agree that the abstract would benefit from including key quantitative results to support the central claims. Although the body of the paper reports reuse ratios, per-phase token counts, and statistical summaries from the five benchmarks, we will revise the abstract to incorporate representative figures (e.g., average KV-cache reuse fractions and phase-transition statistics) so that the claims can be evaluated directly from the abstract. revision: yes

  2. Referee: [Abstract] Abstract (final sentence): the serving implications ('efficient agentic serving must jointly manage repeated model re-entry, persistent context state, and workload-dependent tool behavior') are drawn from ReAct-style traces on five specific benchmarks with Gemma/Qwen models. No analysis or discussion is supplied to establish that these execution graphs, control flows, or context-management strategies are representative of broader agentic workloads (e.g., parallel tool invocation or hierarchical planning), which could materially alter the reported token-reuse and phase-shift statistics.

    Authors: The manuscript is explicitly scoped to ReAct-style agents, as stated in the title, abstract, and methodology. We do not assert that the observed patterns generalize to all agentic paradigms. In the revision we will add an explicit limitations paragraph that delineates the ReAct focus, notes that alternative control flows (parallel invocation, hierarchical planning) could change reuse and phase statistics, and positions the serving implications as guidance for systems targeting ReAct-style workloads. revision: partial

Circularity Check

0 steps flagged

No circularity: purely observational workload measurements

full rationale

The paper reports direct measurements from end-to-end tracing of ReAct-style agents on five benchmarks using Gemma and Qwen models. No equations, fitted parameters, predictions, or derivation steps are present. Claims about token reuse after caching and temporal shifts in tool use are stated as empirical observations from the collected traces, with no reduction to self-defined quantities or self-citation chains. The representativeness assumption is external to any internal derivation and does not create circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an empirical workload-characterization study; the abstract introduces no mathematical derivations, fitted constants, background axioms, or postulated entities.

pith-pipeline@v0.9.1-grok · 5676 in / 1076 out tokens · 23360 ms · 2026-06-29T20:09:32.990545+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 16 canonical work pages · 8 internal anchors

  1. [1]

    Claude Code: Create custom subagents,

    Anthropic, “Claude Code: Create custom subagents, ” https://code. claude.com/docs/en/sub-agents, 2026, accessed: 2026-05-21

    2026

  2. [2]

    Claude Code: Overview,

    ——, “Claude Code: Overview, ” https://code.claude.com/docs/en/ overview, 2026, accessed: 2026-05-21

    2026

  3. [3]

    Efficient and scalable agentic ai with heterogeneous systems,

    Z. Asgar, M. Nguyen, and S. Katti, “Efficient and scalable agentic ai with heterogeneous systems, ”arXiv preprint arXiv:2507.19635, 2025

    work page arXiv 2025

  4. [4]

    Unrolling the Codex agent loop,

    M. Bolin, “Unrolling the Codex agent loop, ” https://openai.com/index/ unrolling-the-codex-agent-loop/, Jan. 2026, openAI. Accessed: 2026- 05-21

    2026

  5. [5]

    arXiv preprint arXiv:2510.09665 , year=

    Y. Cheng, Y. Liu, J. Yao, Y. An, X. Chen, S. Feng, Y. Huang, S. Shen, K. Du, and J. Jiang, “Lmcache: An efficient kv cache layer for enterprise-scale llm inference, ”arXiv preprint arXiv:2510.09665, 2025

    work page arXiv 2025

  6. [6]

    SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?

    X. Deng, J. Da, E. Pan, Y. Y. He, C. Ide, K. Garg, N. Lauffer, A. Park, N. Pasari, C. Rane, K. Sampath, M. Krishnan, S. Kundurthy, S. Hendryx, Z. Wang, V. Bharadwaj, J. Holm, R. Aluri, C. B. C. Zhang, N. Jacobson, B. Liu, and B. Kenstler, “SWE-Bench Pro: Can AI agents solve long-horizon software engineering tasks?”arXiv preprint arXiv:2509.16941, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    Dabstep: Data agent benchmark for multi-step reasoning,

    A. Egg, M. Iglesias Goyanes, F. Kingma, A. Mora, L. von Werra, and T. Wolf, “Dabstep: Data agent benchmark for multi-step reasoning, ” arXiv preprint arXiv:2506.23719, 2025

    work page arXiv 2025

  8. [8]

    AgentQuest: A modular benchmark framework to measure progress and improve LLM agents,

    L. Gioacchini, G. Siracusano, D. Sanvito, K. Gashteovski, D. Friede, R. Bifulco, and C. Lawrence, “AgentQuest: A modular benchmark framework to measure progress and improve LLM agents, ” 2024. [Online]. Available: https://arxiv.org/abs/2404.06411

    work page arXiv 2024

  9. [9]

    Gemma 4: Byte for byte, the most capa- ble open models,

    Google DeepMind, “Gemma 4: Byte for byte, the most capa- ble open models, ” https://blog.google/innovation-and-ai/technology/ developers-tools/gemma-4/, Apr. 2026, accessed: 2026-05-21

    2026

  10. [10]

    Harbor: A framework for evaluating and optimizing agents and models in container environments,

    Harbor Framework Team, “Harbor: A framework for evaluating and optimizing agents and models in container environments, ” Jan. 2026. [Online]. Available: https://github.com/harbor-framework/harbor

    2026

  11. [11]

    Jaeger Documentation,

    Jaeger Authors, “Jaeger Documentation, ” https://www.jaegertracing. io/docs/latest/, 2026, accessed: 2026-05-21

    2026

  12. [12]

    Highly accurate protein structure prediction with alphafold,

    J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. Žídek, A. Potapenkoet al., “Highly accurate protein structure prediction with alphafold, ”Nature, vol. 596, no. 7873, pp. 583–589, 2021

    2021

  13. [13]

    Thunderagent: A simple, fast and program- aware agentic inference system,

    H. Kang, Z. Li, X. Yang, W. Xu, Y. Chen, J. Wang, B. Chen, T. Krishna, C. Xu, and S. Arora, “Thunderagent: A simple, fast and program- aware agentic inference system, ”arXiv preprint arXiv:2602.13692, 2026

    work page arXiv 2026

  14. [14]

    The cost of dynamic reasoning: Demystifying AI agents and test-time scaling from an AI infrastructure perspective,

    J. Kim, B. Shin, J. Chung, and M. Rhu, “The cost of dynamic reasoning: Demystifying AI agents and test-time scaling from an AI infrastructure perspective, ” 2025. [Online]. Available: https://arxiv.org/abs/2506.04301

    work page arXiv 2025

  15. [15]

    Efficient memory management for large language model serving with PagedAttention,

    W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica, “Efficient memory management for large language model serving with PagedAttention, ” inProceedings of the 29th Symposium on Operating Systems Principles (SOSP ’23), 2023, pp. 611–626

    2023

  16. [16]

    Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live

    H. Li, Q. Mang, R. He, Q. Zhang, H. Mao, X. Chen, H. Zhou, A. Cheung, J. Gonzalez, and I. Stoica, “Continuum: Efficient and robust multi-turn LLM agent scheduling with KV cache time-to-live, ” arXiv preprint arXiv:2511.02230, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Parrot: Efficient serving of LLM-based applications with semantic variable,

    C. Lin, Z. Han, C. Zhang, Y. Yang, F. Yang, C. Chen, and L. Qiu, “Parrot: Efficient serving of LLM-based applications with semantic variable, ” in18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24), 2024, pp. 929–945. [Online]. Available: https://www.usenix.org/conference/osdi24/presentation/lin-chaofan

    2024

  18. [18]

    AgentBench: Evaluating LLMs as agents,

    X. Liu, H. Yu, H. Zhang, Y. Xu, X. Lei, H. Lai, Y. Gu, H. Ding, K. Men, K. Yang, S. Zhang, X. Deng, A. Zeng, Z. Du, C. Zhang, S. Shen, T. Zhang, Y. Su, H. Sun, M. Huang, Y. Dong, and J. Tang, “AgentBench: Evaluating LLMs as agents, ” inInternational Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/forum?id=zAdUB0aCTQ

    2024

  19. [19]

    AgentBoard: An analytical evaluation board of multi-turn LLM agents,

    C. Ma, J. Zhang, Z. Zhu, C. Yang, Y. Yang, Y. Jin, Z. Lan, L. Kong, and J. He, “AgentBoard: An analytical evaluation board of multi-turn LLM agents, ” inAdvances in Neural Information Processing Systems, 2024. [Online]. Available: https://openreview.net/forum?id=4S8agvKjle

    2024

  20. [20]

    Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces

    M. A. Merrill, A. G. Shaw, N. Carlini, B. Li, H. Raj, I. Bercovich, L. Shi, J. Y. Shin, T. Walshe, E. K. Buchanan, J. Shen, G. Ye, H. Lin, J. Poulos, M. Wang, M. Nezhurina, J. Jitsev, D. Lu, O. Menis Mastromichalakis, Z. Xu, Z. Chen, Y. Liu, R. Zhang, L. L. Chen, A. Kashyap, J.-L. Uslu, J. Li, J. Wu, M. Yan, S. Bian, V. Sharma, K. Sun, S. Dillmann, A. Ana...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  21. [21]

    GAIA: a benchmark for General AI Assistants

    G. Mialon, C. Fourrier, C. Swift, T. Wolf, Y. LeCun, and T. Scialom, “GAIA: A benchmark for general AI assistants, ”arXiv preprint arXiv:2311.12983, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  22. [22]

    Introducing GPT-5.5,

    OpenAI, “Introducing GPT-5.5, ” https://openai.com/index/ introducing-gpt-5-5/, Apr. 2026, accessed: 2026-05-21

    2026

  23. [23]

    OpenClaw: Personal ai assistant,

    OpenClaw, “OpenClaw: Personal ai assistant, ” https://openclaw.ai/, 2026, accessed: 2026-05-22

    2026

  24. [24]

    OpenTelemetry Documentation,

    OpenTelemetry Authors, “OpenTelemetry Documentation, ” https: //opentelemetry.io/docs/, 2026, accessed: 2026-05-21

    2026

  25. [25]

    The Impact of AI on Developer Productivity: Evidence from GitHub Copilot

    S. Peng, E. Kalliamvakou, P. Cihon, and M. Demirer, “The impact of ai on developer productivity: Evidence from github copilot, ”arXiv preprint arXiv:2302.06590, 2023. [Online]. Available: https://arxiv.org/abs/2302.06590

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    Qwen3.6-27B: Flagship-level coding in a 27b dense model,

    Qwen Team, “Qwen3.6-27B: Flagship-level coding in a 27b dense model, ” https://qwen.ai/blog?id=qwen3.6-27b, Apr. 2026, accessed: 2026-05-21

    2026

  27. [27]

    Towards Understanding, Analyzing, and Optimizing Agentic AI Execution: A CPU-Centric Perspective

    R. Raj, S. Kundu, I. Vohra, H. Wang, and T. Krishna, “Towards understanding, analyzing, and optimizing agentic AI Execution: A CPU-centric perspective, ” 2025. [Online]. Available: https://arxiv.org/abs/2511.00739

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    Alto: An efficient network orchestrator for compound AI systems,

    K. Santhanam, D. Raghavan, M. S. Rahman, T. Venkatesh, N. Kunjal, P. Thaker, P. Levis, and M. Zaharia, “Alto: An efficient network orchestrator for compound AI systems, ” inProceedings of the 4th Workshop on Machine Learning and Systems (EuroMLSys ’24), 2024, pp. 117–125

    2024

  29. [29]

    Toolformer: Language models can teach themselves to use tools,

    T. Schick, J. Dwivedi-Yu, R. Dessi, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom, “Toolformer: Language models can teach themselves to use tools, ” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 36, 2023. [Online]. Available: https://openreview.net/forum?id=Yacmpz84TH

    2023

  30. [30]

    Reflexion: Language agents with verbal reinforcement learning,

    N. Shinn, F. Cassano, B. Labash, A. Gopinath, K. Narasimhan, and S. Yao, “Reflexion: Language agents with verbal reinforcement learning, ” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 36, 2023. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/2023/hash/ 1b44b878bb782e6954cd888628510e90-Abstract-Conference.html 12

    2023

  31. [31]

    ADE-bench: Analytics and data engineering benchmark,

    B. Stancil and dbt Labs, “ADE-bench: Analytics and data engineering benchmark, ” https://github.com/dbt-labs/ade-bench, 2026, accessed: 2026-05-21

    2026

  32. [32]

    Automatic Prefix Caching,

    vLLM Team, “Automatic Prefix Caching, ” https://docs.vllm.ai/en/ latest/features/automatic_prefix_caching/, 2026, vLLM documentation. Accessed: 2026-05-21

    2026

  33. [33]

    Efficient llm serving for agentic work- flows: A data systems perspective,

    N. Wadlom, J. Shen, and Y. Lu, “Efficient llm serving for agentic work- flows: A data systems perspective, ”arXiv preprint arXiv:2603.16104, 2026

    work page arXiv 2026

  34. [34]

    AgentRace: Benchmarking efficiency in LLM agent frameworks,

    Y. Xu, B. Zeng, Z. Qiu, Z. Zhang, G. Yue, X. Liao, H. Jin, and Q. Li, “AgentRace: Benchmarking efficiency in LLM agent frameworks, ” 2026, submitted to ICLR 2026. [Online]. Available: https://openreview.net/forum?id=eUuxWAQA5F

    2026

  35. [35]

    SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

    J. Yang, C. E. Jimenez, A. Wettig, K. Lieret, S. Yao, K. R. Narasimhan, and O. Press, “SWE-agent: Agent-computer interfaces enable automated software engineering, ” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. [Online]. Available: https://arxiv.org/abs/2405.15793

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Webshop: Towards scalable real-world web interaction with grounded language agents,

    S. Yao, H. Chen, J. Yang, and K. Narasimhan, “Webshop: Towards scalable real-world web interaction with grounded language agents, ” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 35, 2022. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/2022/hash/ 82ad13ec01f9fe44c01cb91814fd7b8c-Abstract-Conference.html

    2022

  37. [37]

    React: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, “React: Synergizing reasoning and acting in language models, ” inThe Eleventh International Conference on Learning Representations (ICLR), 2023

    2023

  38. [38]

    KAIROS: Stateful, Context-Aware Power-Efficient Agentic Inference Serving

    Y. Yuan, M. Chowdhury, and N. Talati, “Kairos: Stateful, context- aware power-efficient agentic inference serving, ” 2026. [Online]. Available: https://arxiv.org/abs/2604.16682

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  39. [39]

    arXiv preprint arXiv:2602.09345

    Y. Zheng, J. Fan, Q. Fu, Y. Yang, W. Zhang, and A. Quinn, “AgentCgroup: Understanding and controlling OS resources of AI agents, ” 2026. [Online]. Available: https://arxiv.org/abs/2602.09345

    work page arXiv 2026

  40. [40]

    Language agent tree search unifies reasoning, acting, and planning in language models,

    A. Zhou, K. Yan, M. Shlapentokh-Rothman, H. Wang, and Y.-X. Wang, “Language agent tree search unifies reasoning, acting, and planning in language models, ” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 235. PMLR, 2024, pp. 62 138–62 160. [Online]. Available: https://proceedings....

    2024