Pith. sign in

REVIEW 4 major objections 5 minor 188 references

This paper claims that autonomous-agent capabilities—tool use, file access, and cross-task recall—emerge from the layered integration of a local inference model and a persistent agent runtime, not from the language model alone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 00:51 UTC pith:4NQHIMSV

load-bearing objection The new 15-task benchmark and released artifacts are real, but the central claim of monotonic improvement is largely designed into the protocol rather than demonstrated by it. the 4 major comments →

arxiv 2607.28629 v1 pith:4NQHIMSV submitted 2026-04-16 cs.AI

OpenClaw and Ollama in Agentic AI: Toward Fully Autonomous and Scalable AI Agent Systems

Konstantinos I. Roumeliotis , Ranjan Sapkota This is my paper
classification cs.AI
keywords Agentic AIlayered architecturepersistent memorytool useagent runtimeOllamaOpenClawsystem-level evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that 'agentic' behavior is a property of the full system architecture, not of the language model alone. To support that, it runs a 15-task benchmark in three configurations: a raw local LLM, the same LLM behind an agent runtime with tools but no memory, and the full runtime with persistent on-disk memory. Both of two 4-billion-parameter models follow the same strictly monotonic path in overall task success, from 0.467 to roughly 0.91–0.96 to roughly 0.93–0.98, with perfect memory recall in the full configuration. The authors read this as evidence that each architectural layer contributes independently and that orchestration and memory layers largely erase model-level differences. A reader should care because the result shifts design attention from model size to runtime and memory engineering, and argues for system-level evaluation rather than model-only benchmarks.

Core claim

The central discovery is a strictly monotonic capability ordering across three configurations: raw local LLM only (C1) gives 0.467 success for both models; adding an OpenClaw runtime with tools but wiping memory between tasks (C2) raises it to 0.911 and 0.955; enabling persistent memory (C3) raises it to 0.933 and 0.978. The full-stack configuration achieves memory recall of 1.000 ± 0.000 across all six store–recall tasks, including correct override of a stale stored alias. Identical C1 baselines and near-convergence at C3 indicate an architectural rather than model-specific gradient. Tool-use accuracy dips slightly from C2 to C3, attributed to tool-selection disambiguation.

What carries the argument

The load-bearing comparison is a three-configuration experiment. C1 is Ollama, a local LLM inference server, queried directly with no tools and no memory. C2 routes the same model through OpenClaw, an agent runtime that orchestrates reasoning, tool invocation, and memory, but wipes session files, the SQLite store, and workspace memory after every task. C3 is identical to C2 except that the on-disk memory layer accumulates, enabling long-horizon recall. The 15-task benchmark covers reactive document lookup, tool use (CSV arithmetic, file writing, API calls, cron), and three store–recall pairs plus memory probes. The central identity is the strict ordering C1 < C2 < C3 in overall task success,

Load-bearing premise

The ordering rests on the harness design: under C1 the reactive prompts receive an unresolvable file path while C2/C3 receive the document text, and under C2 memory is wiped between every paired store and recall task, so the measured gradient partly reflects what the experiment intentionally withholds rather than what the runtime adds.

What would settle it

Re-run the 15-task benchmark with file-content enrichment applied to the C1 baseline as well; if reactive-task success rises from 0.000 to 1.000 and C1 overall success approaches or exceeds C2's, the reported C1<C2 jump is an artifact of prompt enrichment. Likewise, if removing the stateless reset in C2 makes T10 recall succeed, then 'persistent' capability is not uniquely attributable to the C3 memory layer.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the ordering holds, agentic capability is an emergent property of the layered stack: the same model behaves reactively alone and agentically once an orchestration layer and persistent memory are added.
  • Different small models converge to similar full-stack performance, so architecture and memory design can matter more than which open-weight model is chosen, at least at this task complexity.
  • Persistent memory makes cross-task recall deterministic in this benchmark (1.000 ± 0.000), including correct overwrite of stale values, implying on-disk memory is a reliable substrate for long-horizon behavior.
  • The measured 11–13x orchestration latency and its partial offset by accumulated memory imply that agentic systems trade latency for capability, and that memory context can reduce reasoning iterations over time.
  • Evaluation of agentic systems should move from model-centric benchmarks to system-level metrics—task success, tool accuracy, memory accuracy, latency, reasoning depth—as the paper's protocol demonstrates.

Where Pith is reading between the lines

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

  • Beyond the paper, the same architecture claim could be stress-tested by keeping C1's prompts identical to C2/C3 (injecting document text) and adding a much larger cloud model as a fourth configuration; if the gradient compresses or reverses, model capability still matters at scale.
  • Beyond the paper, the security discussion implies a testable failure mode: a single adversarial prompt stored in persistent memory could poison all later tasks under C3 but not under C2, which would make memory a liability as well as a capability.
  • Beyond the paper, the task suite is small (15 tasks, 3 repetitions) and the latency comparison mixes tool-bound tasks with memory-bound tasks; a scaled version with longer horizons and open-ended goals would show whether the monotonic ordering persists.
  • Beyond the paper, the slight C3 dip in tool-use accuracy suggests a concrete optimization target: better tool-selection disambiguation in the orchestration layer, which the paper identifies but does not solve.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a layered architecture for Agentic AI in which Ollama provides the LLM inference layer and OpenClaw provides the agent runtime layer (orchestration, tools, memory). After a broad architectural review, it presents a prototype evaluation with three configurations: C1 (Ollama-only), C2 (OpenClaw stateless), and C3 (OpenClaw with persistent memory). The central empirical claim is a strictly monotonic improvement C1 < C2 < C3 in task success rate across two open-weight models on 15 tasks, taken as evidence that autonomous capabilities such as tool use and persistent memory emerge from system-level integration rather than from the model alone.

Significance. If the empirical claim were sound, the paper would provide a useful, reproducible demonstration of how an agent runtime plus persistent memory can be layered over a local LLM. The released code, task definitions, and raw logs are a strength, as is the attempt to compare two different base models. However, the experimental protocol confounds the contribution of the architecture with changes in the input content and with the deliberate wiping of memory in C2. The headline monotonicity is therefore partly encoded in the configuration definitions rather than discovered. The surrounding survey is competent but does not by itself establish the paper's primary conclusion. The tool-use sub-metric also contradicts the claimed 'consistent' improvement, as accuracy drops from C2 to C3 for both models. The central empirical contribution needs a substantially redesigned, properly controlled comparison before it can support the stated conclusions.

major comments (4)
  1. [§4.1, §4.2 (reactive tasks)] C1 receives only the raw relative path string for reactive_notes.md, while C2 and C3 receive the full document text injected by the evaluation harness. The reactive-category jump from 0.000 (C1) to 1.000 (C2/C3) is therefore guaranteed by the input-enrichment policy, not by anything OpenClaw does. A fair baseline for isolating the orchestration layer would provide the same file content to C1 via the prompt. As written, the paper's conclusion that 'reactive capability emerges from the enriched prompt pipeline that the OpenClaw runtime enables' overstates the result: the enrichment is a harness action, and a raw LLM prompted with the pasted text would likely also solve the three tasks.
  2. [§4.2, §4.4 (persistent tasks, C2 reset)] In C2 the harness performs a 'full stateless reset' between each task, so the paired store/recall tasks T9 and T10 cannot both succeed: T10's alias recall is guaranteed to fail regardless of model, runtime, or tool access. The paper itself calls this a 'controlled memory-isolation negative test,' yet it then uses this designed failure to argue that persistent memory is necessary. This is circular for the C2→C3 comparison. The design makes the C2 persistent-category score artificially low, so the observed ordering C1<C2<C3 in the persistent category is not empirical evidence of the architecture's contribution.
  3. [Table 4; §4.4; Finding 1] Tool-use accuracy decreases from C2 to C3 for both models: Qwen3.5:4b goes from 0.944 to 0.833 and Gemma4:e4b from 1.000 to 0.944. This contradicts the abstract's claim of 'performance improving consistently as architectural complexity increases' and Finding 1's 'strict C1<C2<C3 ordering'. The authors attribute the C3 drop to a T11 tool-selection routing issue, but that means the more complex configuration is less reliable on tool use, so the per-category data do not support a strict monotonic architectural benefit. The headline claim should be restricted to the overall success rate and honestly discussed as a trade-off rather than as uniform improvement.
  4. [§4.3, Tables 4–5 (statistical basis)] All conclusions rest on n=3 repetitions, with no significance tests, confidence intervals (beyond means and standard deviations), or effect-size analysis. Given the small sample and the deterministic design failures discussed above, the assertion of a 'strict' ordering and especially the 'zero-variance perfection' for C3 memory recall needs a much stronger statistical and experimental foundation. At minimum, the number of independent runs and task instances must be large enough to distinguish architectural effects from stochastic variation and design artifacts.
minor comments (5)
  1. [Throughout] The paper uses 'OpenClaw' and 'Openclaw' inconsistently. Also, several citations refer to arXiv:260x.xxxxx papers from 2026 that the reader cannot verify; the authors should confirm all references, especially the model documentation links, and include access dates consistently.
  2. [§4.1] The claim that the H100 81GB GPU is used for 4-bit 4B models is relevant, but the paper does not report GPU memory usage, model quantization details, or inference-time parameters beyond temperature and top-p. These details would help reproducibility.
  3. [§4.4] The reactive-category result is described as 'emerging from the enriched prompt pipeline that the OpenClaw runtime enables,' but the harness itself performs the enrichment. The wording should be corrected to avoid attributing a harness action to OpenClaw.
  4. [Figure 8] The figure shows only the overall success rate; given that tool-use accuracy drops from C2 to C3, a figure with per-category error bars and individual task outcomes would be more informative and less likely to be read as supporting uniform monotonic improvement.
  5. [§6] The future-directions discussion is useful but largely speculative and not tied to the experimental results. Consider trimming or grounding it in the specific failure modes identified in the trace logs.

Circularity Check

3 steps flagged

The claimed C1<C2<C3 progression is largely entailed by the configuration definitions: C1 is denied file content and tools, C2 is designed to wipe memory, so the central 'prediction' reduces to the experimental setup.

specific steps
  1. self definitional [Section 4.1, C1 configuration; Section 4.2, Reactive tasks]
    "Local file references in task prompts are passed as-is (the raw relative path string); the evaluation harness applies file-content enrichment only when the OpenClaw orchestration layer is active (C2 and C3). Under C1, the model therefore receives an unresolvable path and cannot access document content."

    The reactive tasks require extracting facts from reactive_notes.md, but C1 is defined so that the model never receives the document text. C2 and C3 are defined to inject the full document text into the prompt. The reactive score of 0.000 in C1 versus 1.000 in C2/C3 is therefore an automatic consequence of the input construction, not a measured capability of OpenClaw. The paper's 'testable prediction' of monotonic reactive improvement is equivalent to the definition of the configurations, so the reactive portion of the central claim is circular.

  2. self definitional [Section 4.1, C2 configuration; Section 4.2, Persistent tasks; Section 4.4]
    "Under C2, the full stateless reset between tasks means no recall task can retrieve information from its paired store task, providing a controlled memory-isolation negative test."

    The persistent tasks are store-recall pairs. C2 is defined to wipe all session, SQLite, and workspace memory between every task, so T10's alias recall cannot succeed regardless of model or runtime. The paper itself later says T10 'fails deterministically' because the reset wipes the entry. C3 is defined to accumulate on-disk memory, making recall possible. The C2<C3 memory gap is thus an artifact of the reset policy, not evidence that only C3 enables memory. The 'controlled memory-isolation negative test' is a design choice that forces the result the paper presents as an empirical finding.

  3. self definitional [Section 4.1, C1 vs C2 configuration; Section 4.4, Tool-use category]
    "Under C1, both models score 0.667 ± 0.000 (4/6 tasks): T4 (total cost), T12 (arithmetic), T11 (API schema), and T13 (cron expression) are solvable through in-context reasoning or training knowledge alone; T5 (CSV item extraction) and T6 (file write and verify) require tool access and are structurally unsolvable without it."

    C1 is defined as having 'no tool availability,' while C2 and C3 make 'the full toolset available.' Since T5 requires reading a CSV and T6 requires writing a file, both are impossible under C1 by construction. The tool-use improvement from 0.667 in C1 to 0.944/1.000 in C2 is therefore partly forced by the harness's tool-access configuration, not by an independently demonstrated emergent orchestration capability. The monotonicity claim across the tool category is again built into the definition of which tools each configuration may use.

full rationale

The paper's central empirical claim is that task success increases strictly from the Ollama-only baseline (C1), through the stateless OpenClaw configuration (C2), to the persistent full-stack configuration (C3), and that this demonstrates capabilities 'emerge from system-level integration rather than standalone models.' The reduction, however, is visible in the paper's own configuration definitions. For reactive tasks, C1 is defined to receive only an unresolvable path while C2/C3 receive the document text via harness injection, guaranteeing the reactive 0-to-1 jump. For persistent tasks, C2 is defined with a full stateless reset between tasks, guaranteeing that recall tasks fail, while C3 is defined with accumulating memory, guaranteeing recall is possible. For tool-use tasks, C1 has no tools while C2/C3 do, so the two tool-requiring tasks are structurally unsolvable in C1. These three design choices together force most of the C1<C2<C3 progression; the paper even labels the C2 persistent failure a 'controlled memory-isolation negative test' and then uses that designed failure as evidence for the value of persistent memory. This is not a case of fitted parameters being renamed predictions, nor is there a load-bearing self-citation chain or imported uniqueness theorem; the paper is transparent and releases its code and data. But the central empirical demonstration is substantially circular because the outcomes that are said to confirm the architecture are entailed by the way the configurations were constructed. The score is 7 rather than higher because some components, such as tool-selection failures and latency, are not fully forced by the definitions, and the paper does not misrepresent its negative-test labeling. The core issue is that the headline prediction reduces to the experimental setup, so the claim of emergent capability from system integration is not independently tested.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No numeric free parameters are fitted; the load-bearing choices are the experimental protocol itself. The empirical claim rides on ad hoc assumptions about how to isolate layers, and those assumptions are exactly what make the headline result come out. The entities (OpenClaw, Ollama, the two models) are existing systems, not invented by this paper.

axioms (4)
  • ad hoc to paper C1's lack of file-content enrichment isolates the LLM inference layer; providing the same content via the harness in C2/C3 is a fair measure of runtime contribution.
    This assumption is load-bearing: without it, the reactive-task gap (0→1) says nothing about OpenClaw. It is introduced in §4.1 C1/C2 definitions.
  • ad hoc to paper Wiping the memory store between paired store/recall tasks in C2 is a valid way to show that persistent memory is necessary.
    The T9→T10 failure under C2 is predetermined by the stateless reset; the test assumes that memory persistence is the only relevant architectural difference between C2 and C3, while other state is also reset.
  • domain assumption Two models times three repetitions is sufficient to generalize about architecture.
    No power analysis, significance tests, or per-category error bars are provided; §4.4 still makes strong claims about determinism and architecture-driven performance.
  • domain assumption OpenClaw version 2026.4.5 and the named model identifiers (Qwen3.5:4b, Gemma4:e4b) exist and behave as described.
    The paper cites model URLs and an OpenClaw version but provides no verification; the reviewer could not independently confirm these artifacts.

pith-pipeline@v1.3.0-alltime-deepseek · 31985 in / 12529 out tokens · 110100 ms · 2026-08-03T00:51:36.742391+00:00 · methodology

0 comments
read the original abstract

The rapid transition from reactive large language models (LLMs) to persistent, action-capable systems has exposed critical gaps in the architectural understanding of Agentic AI, particularly in separating inference, orchestration, and execution layers for autonomous AI agents. Despite recent advances, unified frameworks for designing and evaluating full-stack agentic systems remain limited. This paper presents a comprehensive, layered architecture for Agentic AI, outlining the evolution from reactive LLM interfaces to persistent, goal-driven autonomous AI agents with memory, planning, and continuous execution. We analyze OpenClaw and Ollama as a full-stack Agentic AI system, where Ollama serves as the LLM inference layer and OpenClaw enables agent runtime orchestration, integrating reasoning, tool use, and action execution. A prototype experimental validation of the OpenClaw-Ollama architecture demonstrates that capabilities such as persistent memory, tool utilization, and adaptive decision-making emerge from system-level integration rather than standalone models, with performance improving consistently as architectural complexity increases. The study further examines challenges in scalability, security, privacy, governance, and evaluation of agentic systems, highlighting the need for robust benchmarking and system-level design. Future directions include scalable multi-agent architectures, distributed autonomous systems, and human-aware Agentic AI frameworks for responsible deployment. Overall, this work establishes a unified architectural foundation for Agentic AI, validates the effectiveness of full-stack autonomous AI agents, and provides a roadmap for building scalable, secure, and trustworthy agentic systems. All models, code, and datasets are publicly released to support reproducibility and benchmarking.

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

188 extracted references · 23 linked inside Pith

  1. [1]

    Computers, Materials & Continua80(2) (2024)

    Chen, Z., Xu, L., Zheng, H., Chen, L., Tolba, A., Zhao, L., Yu, K., Feng, H.: Evolution and prospects of foundation models: From large language models to large multimodal models. Computers, Materials & Continua80(2) (2024)

    2024

  2. [2]

    Annual Reviews in Control60, 101021 (2025)

    Sarikaya, R.: Path to artificial general intelligence: past, present, and future. Annual Reviews in Control60, 101021 (2025)

    2025

  3. [3]

    ACM Computing Surveys (2023)

    Liu, X., Wang, J., Yuan, X., Sun, J., Dong, G., Di, P., Wang, W., Wang, D.: Prompting frameworks for large language models: A survey. ACM Computing Surveys (2023)

    2023

  4. [4]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2026)

    Zheng, J., Shi, C., Cai, X., Li, Q., Zhang, D., Li, C., Yu, D., Ma, Q.: Lifelong learning of large language model based agents: A roadmap. IEEE Transactions on Pattern Analysis and Machine Intelligence (2026)

    2026

  5. [5]

    agentic ai: A conceptual taxonomy, applications and challenges

    Sapkota, R., Roumeliotis, K.I., Karkee, M.: Ai agents vs. agentic ai: A conceptual taxonomy, applications and challenges. Information Fusion, 103599 (2025)

    2025

  6. [6]

    Artificial Intelligence Review 59(1), 11 (2025)

    Abou Ali, M., Dornaika, F., Charafeddine, J.: Agentic ai: a comprehensive survey of architectures, applications, and future directions. Artificial Intelligence Review 59(1), 11 (2025)

    2025

  7. [7]

    Artificial Intelligence Review (2026)

    Alva, L., Pandey, B.: Agentic ai systems in the age of generative models: architec- tures, cloud scalability, and real-world applications. Artificial Intelligence Review (2026)

    2026

  8. [8]

    Future Internet17(9), 404 (2025)

    Bandi, A., Kongari, B., Naguru, R., Pasnoor, S., Vilipala, S.V.: The rise of agentic ai: A review of definitions, frameworks, architectures, applications, evaluation metrics, and challenges. Future Internet17(9), 404 (2025)

    2025

  9. [9]

    Information & Management, 104317 (2026)

    Kumar, N., Wei, X., Zhang, H.: Agentic artificial intelligence as a new frontier in information systems: Promise, peril, and research opportunities. Information & Management, 104317 (2026)

    2026

  10. [10]

    arXiv preprint arXiv:2601.12538 (2026)

    Wei, T., Li, T.-W., Liu, Z., Ning, X., Yang, Z., Zou, J., Zeng, Z., Qiu, R., Lin, X., Fu, D., et al.: Agentic reasoning for large language models. arXiv preprint arXiv:2601.12538 (2026)

    Pith/arXiv arXiv 2026

  11. [11]

    F1000Research14(905), 905 (2025)

    Adabara, I., Sadiq, B.O., Shuaibu, A.N., Danjuma, Y.I., Maninti, V.: Trustworthy agentic ai systems: a cross-layer review of architectures, threat models, and 36 governance strategies for real-world deployment. F1000Research14(905), 905 (2025)

    2025

  12. [12]

    Vicinagearth1(1), 9 (2024)

    Li, X., Wang, S., Zeng, S., Wu, Y., Yang, Y.: A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges. Vicinagearth1(1), 9 (2024)

    2024

  13. [13]

    ACM Transactions on Software Engineering and Methodology34(5), 1–30 (2025)

    He,J.,Treude,C.,Lo,D.:Llm-basedmulti-agentsystemsforsoftwareengineering: Literature review, vision, and the road ahead. ACM Transactions on Software Engineering and Methodology34(5), 1–30 (2025)

    2025

  14. [14]

    Universitat Oberta de Catalunya (UOC) (2025)

    Rodríguez Quiñones, A.: Privacy-focused llm for local data processing: Imple- menting ollama and rag to securely query personal files in closed environments. Universitat Oberta de Catalunya (UOC) (2025)

    2025

  15. [15]

    In: Natural Language Analytics with Generative Large-Language Models: A Practical Approach with Ollama and Open-Source LLMs, pp

    Marcondes, F.S., Gala, A., Magalhães, R., Britto, F., Durães, D., Novais, P.: Using ollama. In: Natural Language Analytics with Generative Large-Language Models: A Practical Approach with Ollama and Open-Source LLMs, pp. 23–35. Springer, ??? (2025)

    2025

  16. [16]

    arXiv preprint arXiv:2602.19810 (2026)

    Weidener, L., Brkić, M., Jovanović, M., Singh, R., Ulgac, E., Meduri, A.: Open- claw, moltbook, and clawdlab: From agent-only social networks to autonomous scientific research. arXiv preprint arXiv:2602.19810 (2026)

    arXiv 2026

  17. [17]

    arXiv preprint arXiv:2603.10092 (2026)

    Borjigin, A., Stadnyk, I., Bilski, B., Hovorov, S., Pidturkina, S.: Execution is the new attack surface: Survivability-aware agentic crypto trading with openclaw- style local executors. arXiv preprint arXiv:2603.10092 (2026)

    arXiv 2026

  18. [18]

    arXiv preprint arXiv:2603.12644 (2026)

    Ying, Z., Yang, X., Wu, S., Song, Y., Qu, Y., Li, H., Li, T., Wang, J., Liu, A., Liu, X.: Uncovering security threats and architecting defenses in autonomous agents: A case study of openclaw. arXiv preprint arXiv:2603.12644 (2026)

    arXiv 2026

  19. [19]

    arXiv preprint arXiv:2603.11619 (2026)

    Deng, X., Zhang, Y., Wu, J., Bai, J., Yi, S., Zou, Z., Xiao, Y., Qiu, R., Ma, J., Chen, J., et al.: Taming openclaw: Security analysis and mitigation of autonomous llm agent threats. arXiv preprint arXiv:2603.11619 (2026)

    arXiv 2026

  20. [20]

    arXiv preprint arXiv:2601.12560 (2026)

    Buyya, R., et al.: Agentic artificial intelligence (ai): Architectures, taxonomies, and evaluation of large language model agents. arXiv preprint arXiv:2601.12560 (2026)

    arXiv 2026

  21. [21]

    arXiv preprint arXiv:2603.14312 (2026)

    Wang, F.Y., Marom, L., Pal, S., Luu, R.K., Lu, W., Berkovich, J.A., Buehler, M.J.: Autonomous agents coordinating distributed discovery through emergent artifact exchange. arXiv preprint arXiv:2603.14312 (2026)

    arXiv 2026

  22. [22]

    EarthArXiv (2026)

    Zhang, Z., Valeo, C.: Agentic modelling pipeline: Reproducible rapid stormwater modelling management system with openclaw. EarthArXiv (2026)

    2026

  23. [23]

    arXiv preprint arXiv:2603.08425 37 (2026)

    Mo, X.: Ironengine: Towards general ai assistant. arXiv preprint arXiv:2603.08425 37 (2026)

    arXiv 2026

  24. [24]

    Authorea Preprints (2026)

    Flynn, D.C.: Cognitive integrity security: Securing human-ai interaction in large language model systems. Authorea Preprints (2026)

    2026

  25. [25]

    arXiv preprint arXiv:2603.16572 (2026)

    Holzbauer, F., Schmidt, D., Gegenhuber, G., Schrittwieser, S., Ullrich, J.: Mali- cious or not: Adding repository context to agent skill classification. arXiv preprint arXiv:2603.16572 (2026)

    Pith/arXiv arXiv 2026

  26. [26]

    arXiv preprint arXiv:2603.25100 (2026)

    Ruan, A.: From logic monopoly to social contract: Separation of power and the institutional foundations for autonomous agent economies. arXiv preprint arXiv:2603.25100 (2026)

    arXiv 2026

  27. [27]

    Journal of Computer Science and Technology Studies7(5), 182–193 (2025)

    Garg, V.: Designing the mind: How agentic frameworks are shaping the future of ai behavior. Journal of Computer Science and Technology Studies7(5), 182–193 (2025)

    2025

  28. [28]

    AI6(6), 123 (2025)

    Olujimi, P.A., Owolawi, P.A., Mogase, R.C., Wyk, E.V.: Agentic ai frameworks in smmes: a systematic literature review of ecosystemic interconnected agents. AI6(6), 123 (2025)

    2025

  29. [29]

    Psychology & Marketing43(3), 538–555 (2026)

    Frank, D.-A., Folwarczny, M., Otterbring, T.: Consumer acceptance of high- autonomy ai assistants is driven by perceived benefits in online shopping settings characterized by scarcity. Psychology & Marketing43(3), 538–555 (2026)

    2026

  30. [30]

    International Journal of Innovative Science and Research Technology, 1502–1512 (2025)

    Kumar, G.: Agentic ai in finance: Building autonomous financial assistants on aws for personalized customer solutions. International Journal of Innovative Science and Research Technology, 1502–1512 (2025)

    2025

  31. [31]

    ai: Does ai learning assistant enhance students’ innovation behavior? Education and Information Technologies 30(12), 17483–17530 (2025)

    Luo, L., Hu, J., Zheng, Y., Li, C.: Human vs. ai: Does ai learning assistant enhance students’ innovation behavior? Education and Information Technologies 30(12), 17483–17530 (2025)

    2025

  32. [32]

    Journal of Digital Technologies and Law3(3), 431–445 (2025)

    Bowen, G.: Agentic artificial intelligence: Legal and ethical challenges of autonomous systems. Journal of Digital Technologies and Law3(3), 431–445 (2025)

    2025

  33. [33]

    Business & Information Systems Engineering, 1–12 (2025)

    Piccoli, G., Rodriguez, J., Mahmoud, A.: Large action models for programmatic orchestration. Business & Information Systems Engineering, 1–12 (2025)

    2025

  34. [34]

    European Index Library of European International Journal of Multidisciplinary Research and Management Studies6(01), 108–113 (2026)

    Nygaard, L.H.: Agentic artificial intelligence orchestration and interoperable multi-agent frameworks in enterprise commerce transformation. European Index Library of European International Journal of Multidisciplinary Research and Management Studies6(01), 108–113 (2026)

    2026

  35. [35]

    In: Proceedings of the 42nd International Conference 38 on Machine Learning, vol

    Chopra, A., Bhattacharya, S., Leibo, J.Z., Raskar, R.: Levels of social orchestra- tion for agentic systems. In: Proceedings of the 42nd International Conference 38 on Machine Learning, vol. 267 (2025)

    2025

  36. [36]

    ACM Transactions on Intelligent Systems and Technology (2025)

    Park, S., Jeon, S., Lee, C., Jeon, S., Kim, B.-S., Lee, J.: A survey on inference engines for large language models: Perspectives on optimization and efficiency. ACM Transactions on Intelligent Systems and Technology (2025)

    2025

  37. [37]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Gao, Z., Xu, Y., Thebault-Spieker, J.: Localbench: Benchmarking llms on county- level local knowledge and reasoning. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, pp. 38487–38495 (2026)

    2026

  38. [38]

    SoftwareX30, 102183 (2025)

    Vake, D., Vičič, J., Tošić, A.: Hive: A secure, scalable framework for distributed ollama inference. SoftwareX30, 102183 (2025)

    2025

  39. [39]

    In: Proceedings of the 31st International Conference on Computational Linguistics, pp

    Li, X.: A review of prominent paradigms for llm-based agents: Tool use, planning (including rag), and feedback learning. In: Proceedings of the 31st International Conference on Computational Linguistics, pp. 9760–9779 (2025)

    2025

  40. [40]

    In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp

    Hao, S., Gu, Y., Ma, H., Hong, J., Wang, Z., Wang, D., Hu, Z.: Reasoning with language model is planning with world model. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 8154–8173 (2023)

    2023

  41. [41]

    arXiv preprint arXiv:2510.01833 (2025)

    Dou, Z., Zhao, Q., Wan, Z., Zhang, D., Wang, W., Raiyan, T., Chen, B., Pan, Q., Ouyang, Y., Gao, Z., et al.: Plan then action: High-level planning guidance reinforcement learning for llm reasoning. arXiv preprint arXiv:2510.01833 (2025)

    Pith/arXiv arXiv 2025

  42. [42]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp

    Salama, R., Cai, J., Yuan, M., Currey, A., Sunkara, M., Zhang, Y., Benajiba, Y.: Meminsight: Autonomous memory augmentation for llm agents. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 33124–33140 (2025)

    2025

  43. [43]

    In: Findings of the Association for Computational Linguistics: ACL 2025, pp

    Xu, M., Liang, G., Chen, K., Wang, W., Zhou, X., Yang, M., Zhao, T., Zhang, M.: Memory-augmented query reconstruction for llm-based knowledge graph reasoning. In: Findings of the Association for Computational Linguistics: ACL 2025, pp. 24068–24084 (2025)

    2025

  44. [44]

    Frontiers in Psychology16, 1591618 (2025)

    Hong, C., He, Q.: Enhancing memory retrieval in generative agents through llm-trained cross attention networks. Frontiers in Psychology16, 1591618 (2025)

    2025

  45. [45]

    In: Companion Proceedings of the ACM on Web Conference 2025, pp

    Wang, C., Zhang, Y., Zhu, F., Zhang, J., Shi, T., Feng, F.: Leveraging mem- ory retrieval to enhance llm-based generative recommendation. In: Companion Proceedings of the ACM on Web Conference 2025, pp. 1346–1350 (2025)

    2025

  46. [46]

    IEEE Transactions on Cognitive Communications and Networking (2025) 39

    Luo, H., Liu, Y., Zhang, R., Wang, J., Sun, G., Niyato, D., Yu, H., Xiong, Z., Wang, X., Shen, X.: Toward edge general intelligence with multiple-large language model (multi-llm): architecture, trust, and orchestration. IEEE Transactions on Cognitive Communications and Networking (2025) 39

    2025

  47. [47]

    In: 2025 International Conference on Computer and Applications (ICCA), pp

    Rudra, A., Agrawal, M.: Composable ai stack for intelligent agents: Modular orchestration using context routing, memory, and tools. In: 2025 International Conference on Computer and Applications (ICCA), pp. 1–6 (2025). IEEE

    2025

  48. [48]

    Cryptology ePrint Archive (2026)

    Andreoletti, D., Rudi, A., Carpanzano, E., Leidi, T.: Privacy-preserving llm infer- ence in practice: A comparative survey of techniques, trade-offs, and deployability. Cryptology ePrint Archive (2026)

    2026

  49. [49]

    In: Proceedings of the Nineteenth ACM International Conference on Web Search and Data Mining, pp

    Zhang, K., Wang, C., Peng, L., Go, A., Liu, X.: Privacy-preserved llm cas- cade via cot-enhanced policy learning. In: Proceedings of the Nineteenth ACM International Conference on Web Search and Data Mining, pp. 955–964 (2026)

    2026

  50. [50]

    In: Proceedings of the SC’25 Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis, pp

    Tanikanti, A., Côté, B., Guo, Y., Chen, L., Saint, N., Chard, R., Raffenetti, K., Thakur, R., Uram, T., Foster, I.,et al.: First: Federated inference resource scheduling toolkit for scientific ai model access. In: Proceedings of the SC’25 Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis, pp. 52–60 (2025)

    2025

  51. [51]

    Automation in Construction183, 106805 (2026)

    Gopee, M.A., Prieto, S.A., Soto, B.G.: Self-hosted multimodal large language models for speech-driven perception and navigation in construction robotics. Automation in Construction183, 106805 (2026)

    2026

  52. [52]

    arXiv preprint arXiv:2601.09527 (2026)

    Knoop, J., Holtmann, H.: Private llm inference on consumer blackwell gpus: A practical guide for cost-effective local deployment in smes. arXiv preprint arXiv:2601.09527 (2026)

    arXiv 2026

  53. [53]

    IEEE Transactions on Neural Networks and Learning Systems (2025)

    Zhao, Z., Fu, Y., Chai, J., Zhu, Y., Zhao, D.: Meta learning task representation in multiagent reinforcement learning: From global inference to local inference. IEEE Transactions on Neural Networks and Learning Systems (2025)

    2025

  54. [54]

    arXiv preprint arXiv:2602.13692 (2026)

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

    Pith/arXiv arXiv 2026

  55. [55]

    arXiv preprint arXiv:2603.10387 (2026)

    Shan, Z., Xin, J., Zhang, Y., Xu, M.: Don’t let the claw grip your hand: A security analysis and defense framework for openclaw. arXiv preprint arXiv:2603.10387 (2026)

    arXiv 2026

  56. [56]

    arXiv preprint arXiv:2603.27517 (2026)

    Suwansathit, S., Zhang, Y., Gu, G.: A systematic taxonomy of security vulner- abilities in the openclaw ai agent framework. arXiv preprint arXiv:2603.27517 (2026)

    Pith/arXiv arXiv 2026

  57. [57]

    In: Intelligent Language Services: Theory and Practice with Large Language Models, pp

    Yu, J.S., Yao, Y.: Agents and workflows. In: Intelligent Language Services: Theory and Practice with Large Language Models, pp. 113–141. Springer, ??? (2026)

    2026

  58. [58]

    Sustainable Development (2026) 40

    Lee, Y., Park, E.: Toward sustainable agentic ai systems: A survey of architectures and methodologies. Sustainable Development (2026) 40

    2026

  59. [59]

    In: Proceedings of the 16th ACM SIGOPS Asia-Pacific Workshop on Systems, pp

    Cornacchia, A., Alabdulaal, I., Saghier, I., Mirdad, A., Fayoumi, O., Canini, M.: Between promise and pain: The reality of automating failure analysis in microservices with llms. In: Proceedings of the 16th ACM SIGOPS Asia-Pacific Workshop on Systems, pp. 155–167 (2025)

    2025

  60. [60]

    ACM Computing Surveys58(6), 1–36 (2025)

    He, F., Zhu, T., Ye, D., Liu, B., Zhou, W., Yu, P.S.: The emerged security and privacy of llm agent: A survey with case studies. ACM Computing Surveys58(6), 1–36 (2025)

    2025

  61. [61]

    Information17(1), 54 (2026)

    Gulyamov, S., Gulyamov, S., Rodionov, A., Khursanov, R., Mekhmonov, K., Babaev, D., Rakhimjonov, A.: Prompt injection attacks in large language models and ai agent systems: A comprehensive review of vulnerabilities, attack vectors, and defense mechanisms. Information17(1), 54 (2026)

    2026

  62. [62]

    International Journal of Computer & Organization Trends 12(1), 3–10 (2022)

    Thaiya, M.S., Julia, K., Mbugua, S.: On software modular architecture: Concepts, metrics and trends. International Journal of Computer & Organization Trends 12(1), 3–10 (2022)

    2022

  63. [63]

    In: 2003 Proceedings 29th Euromicro Conference, pp

    Mari, Eila: The impact of maintainability on component-based software systems. In: 2003 Proceedings 29th Euromicro Conference, pp. 25–32 (2003). IEEE

    2003

  64. [64]

    IEEE Communications Surveys & Tutorials28, 4285–4318 (2026)

    Zhang, R., Liu, G., Liu, Y., Zhao, C., Wang, J., Xu, Y., Niyato, D., Kang, J., Li, Y., Mao, S.,et al.: Toward edge general intelligence with agentic ai and agen- tification: Concepts, technologies, and future directions. IEEE Communications Surveys & Tutorials28, 4285–4318 (2026)

    2026

  65. [65]

    Algorithms18(8), 499 (2025)

    Brohi, S., Mastoi, Q.-u.-a., Jhanjhi, N., Pillai, T.R.: A research landscape of agentic ai and large language models: Applications, challenges and future directions. Algorithms18(8), 499 (2025)

    2025

  66. [66]

    AGPH Books| AG Publishing House, ??? (2025)

    Rohitha, K.P.: Agentic AI-Building Autonomous Intelligent Systems for Enter- prise Applications. AGPH Books| AG Publishing House, ??? (2025)

    2025

  67. [67]

    International Journal of Applied Resilience and Sustainability2(2), 142–167 (2026)

    Biswas, B., Sarkar, S.: Responsible agentic artificial intelligence governance: Risk, safety, and ethical challenges in autonomous systems. International Journal of Applied Resilience and Sustainability2(2), 142–167 (2026)

    2026

  68. [68]

    Synergia: A Journal of Multidisciplinary Innovation7(7) (2025)

    Janakiraman, A.: Governance and accountability frameworks for ai agents. Synergia: A Journal of Multidisciplinary Innovation7(7) (2025)

    2025

  69. [69]

    In: 2025 5th International Conference on AI-ML-Systems (AIMLSystems), pp

    Sarin, S., Singh, L., Sarmah, B., Mehta, D.: Memoria: A scalable agentic mem- ory framework for personalized conversational ai. In: 2025 5th International Conference on AI-ML-Systems (AIMLSystems), pp. 32–39 (2025). IEEE

    2025

  70. [70]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp

    Kang, J., Ji, M., Zhao, Z., Bai, T.: Memory os of ai agent. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 25972–25981 (2025) 41

    2025

  71. [71]

    Autonomous agents and multi-agent systems1(1), 89–111 (1998)

    Lesser, V.R.: Reflections on the nature of multi-agent coordination and its implications for an agent architecture. Autonomous agents and multi-agent systems1(1), 89–111 (1998)

    1998

  72. [72]

    Computing 107(1), 2 (2025)

    Ahmed, I., Syed, M.A., Maaruf, M., Khalid, M.: Distributed computing in multi- agent systems: a survey of decentralized machine learning approaches. Computing 107(1), 2 (2025)

    2025

  73. [73]

    Electronics15(2), 385 (2026)

    Álvarez-López, C., González-Briones, A., Li, T.: Explainable ai and multi-agent systems for energy management in iot-edge environments: A state of the art review. Electronics15(2), 385 (2026)

    2026

  74. [74]

    Fundação Universidade Federal de Mato Grosso do Sul (2025)

    Coelho, R.M., Lima, V.M.A.: Adaptive load balancing for distributed llm infer- ence using ollama. Fundação Universidade Federal de Mato Grosso do Sul (2025)

    2025

  75. [75]

    In: International Conference on Human-Computer Interaction, pp

    Liu, Z., Wang, Q., Lu, H., Wang, Y.: Feasibility and usability practice on local hosting open source large language models (llms) including llama 3.2 vision 90b in multi-functional agentic artificial intelligence (ai) system to drive service for design in the latest affordable small personal computer (pc) system. In: International Conference on Human-Comp...

    2025

  76. [76]

    Chen, S., Wang, Q., Yu, G., Wang, X., Zhu, L.: Clawed and dangerous: Can we trust open agentic systems? arXiv preprint arXiv:2603.26221 (2026)

    arXiv 2026

  77. [77]

    arXiv preprint arXiv:2603.10165 (2026)

    Wang, Y., Chen, X., Jin, X., Wang, M., Yang, L.: Openclaw-rl: Train any agent simply by talking. arXiv preprint arXiv:2603.10165 (2026)

    Pith/arXiv arXiv 2026

  78. [78]

    arXiv preprint arXiv:2508.11126 (2025)

    Wang, H., Gong, J., Zhang, H., Xu, J., Wang, Z.: Ai agentic programming: A sur- vey of techniques, challenges, and opportunities. arXiv preprint arXiv:2508.11126 (2025)

    arXiv 2025

  79. [79]

    arXiv preprint arXiv:2511.18538 (2025)

    Yang, J., Liu, X., Lv, W., Deng, K., Guo, S., Jing, L., Li, Y., Liu, S., Luo, X., Luo, Y., et al.: From code foundation models to agents and applications: A comprehensive survey and practical guide to code intelligence. arXiv preprint arXiv:2511.18538 (2025)

    arXiv 2025

  80. [80]

    In: 2025 ACM/IEEE 28th International Conference on Model Driven Engineering Languages and Systems Companion (MODELS-C), pp

    Hermawan, R., Bucaioni, A., Enoiu, E., Afzal, W.: Benchmarking large lan- guage models for root cause analysis in train control software testing. In: 2025 ACM/IEEE 28th International Conference on Model Driven Engineering Languages and Systems Companion (MODELS-C), pp. 648–657 (2025). IEEE

    2025

Showing first 80 references.