pith. sign in

arxiv: 2606.20631 · v1 · pith:WVNHNQEWnew · submitted 2026-05-29 · 💻 cs.AI · cs.LG

Harnessing Agent Skills: Architectural Patterns and a Reference Architecture for Skill-Mediated LLM Agents

Boming Xia , Liming Zhu , Zhenchang Xing , Qinghua Lu , Dino Sejdinovic , Xiwei Xu This is my paper

Pith reviewed 2026-06-28 22:38 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords LLM agentsagent skillsarchitectural patternsreference architectureskill harnessingskill-in-useAI agent designsoftware architecture
0
0 comments X

The pith

Ten architectural patterns and a four-layer reference architecture organize skill harnessing responsibilities in LLM agent systems.

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

The paper studies how LLM agents transition from static skill artefacts to dynamic skill-in-use, where each artefact is selected, bound to constraints, interpreted stochastically, and recorded as evidence. It catalogues ten empirically grounded patterns, five core and five supporting, and synthesizes them into a reference architecture with four responsibility layers: Supply Chain, Mediation, Execution Control, and Evidence & Feedback. A sympathetic reader would care because the patterns and architecture supply a shared vocabulary and diagnostic frame for examining how agent systems manage reusable behavioral knowledge and its consequences. The patterns were identified from existing systems and validated by cross-instantiation on eight selected systems.

Core claim

Agent skill harnessing consists of the architectural responsibilities that govern the transition from persistent skill artefacts to skill-in-use, bound the executable consequences, and capture evidence for attribution, verification, repair, and evolution. The paper presents a catalogue of ten empirically grounded architectural patterns and synthesises them into a reference architecture with four responsibility layers that together provide a vocabulary and diagnostic frame for analysing skill-harnessing responsibilities across agent systems.

What carries the argument

The skill-in-use relation together with the ten architectural patterns (five core, five supporting) that structure responsibilities for selection, binding, constraint enforcement, interpretation, and evidence recording.

If this is right

  • Agent systems can be diagnosed for completeness across the four responsibility layers.
  • The patterns supply reusable solutions for common challenges in skill discovery, activation, and constraint binding.
  • Cross-system comparison and analysis become possible using a shared reference structure.
  • Structured evidence recording supports attribution, verification, and iterative improvement of skills.

Where Pith is reading between the lines

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

  • The four layers could be used as an audit checklist when reviewing codebases of existing LLM agent frameworks.
  • Similar pattern catalogues might be developed for other reusable artefacts such as memory modules or planning templates.
  • Interoperable skill repositories could emerge if multiple platforms adopt the same reference layers.

Load-bearing premise

The ten patterns identified from the eight selected systems generalize as a diagnostic frame for skill-harnessing responsibilities in arbitrary LLM agent architectures.

What would settle it

An LLM agent system that uses reusable skill artefacts but whose responsibilities for selection, binding, execution control, and evidence recording do not map onto the ten patterns or four layers.

Figures

Figures reproduced from arXiv: 2606.20631 by Boming Xia, Dino Sejdinovic, Liming Zhu, Qinghua Lu, Xiwei Xu, Zhenchang Xing.

Figure 1
Figure 1. Figure 1: Conceptual boundary of agent skill harnessing: skill artefacts are mediated into run-scoped skill-in-use, materialised as active skill context within the agent runtime, and linked to evidence and candidate changes from skill-mediated behaviour. the static skill artefact but the stochastic interpretation and enactment of the artefact by an LLM agent. No individual property of a skill artefact is architectur… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of research methodology. Run-to-artefact feedback. Evidence from agent runs may suggest that existing skills should be revised or that new skills should be created. What makes this concern skill￾specific is that the feedback target is a shared behavioural guidance artefact that is reused across future runs and, potentially, by other agents. Local fixes that overfit a single run can therefore propa… view at source ↗
Figure 3
Figure 3. Figure 3: Progressive Skill Activation: skill artefacts become selectable through descriptors, and activation constructs skill￾in-use for the agent runtime. Core patterns respond directly to the five architectural properties of skill-in-use identified in Section 2.3. Pro￾gressive Skill Activation addresses selective runtime par￾ticipation; Skill–Execution Authority Separation addresses capability reference without a… view at source ↗
Figure 5
Figure 5. Figure 5: Verifiable Skill Contract: skill-scoped criteria are checked against skill-use evidence by an independent verifier. can carry Codex-specific metadata for invocation policy and tool dependencies. Verification-conditioned authority is supported as a research-informed refinement by work on verifiable skill artefacts and side-effect-sensitive capability gating (Metere, 2026). 4.3. Verifiable Skill Contract Int… view at source ↗
Figure 6
Figure 6. Figure 6: Runtime Skill-BOM: a per-run, skill-centred prove￾nance manifest linking skill artefacts, participation decisions, and run evidence. Forces. (i) Run specificity vs. catalogue overbreadth. A repository-level inventory over-approximates any run. Attri￾bution requires the smaller set of skill artefacts and decisions actually associated with that run. (ii) Attribution fidelity vs. evidence overhead. High-fidel… view at source ↗
Figure 7
Figure 7. Figure 7: Skill–Agent Co-Evolution Loop: accepted skills shape runs, and run evidence produces validated skill changes for future runs. Variants. Architectural variants differ mainly in where run￾derived skill changes are made reusable. In run-local evolu￾tion, a session, project, or workspace turns its own traces, conversations, failures, or successful procedures into skill changes for later local reuse. In reposit… view at source ↗
Figure 8
Figure 8. Figure 8: Pattern-oriented reference architecture for skill harnessing. The architecture organises responsibilities across Supply Chain, Mediation, Execution Control, and Evidence & Feedback layers, with Policy/Configuration and Identifier as cross-cutting substrates. Solid arrows denote operational, artefact, request, or control flow; dotted arrows denote evidence or provenance flow. observes and acts on the run, p… view at source ↗
read the original abstract

Agent skills externalise reusable agent-facing behavioural knowledge and guidance as persistent artefacts that can be discovered, activated, and interpreted by LLM agents. Although a skill artefact is static at rest, its architectural responsibilities arise in use, when the artefact is selected for a run, bound to context and authority constraints, interpreted by a stochastic agent, and recorded as run evidence. We call this run-specific relation skill-in-use. This paper studies agent skill harnessing: the architectural responsibilities that govern the transition from skill artefacts to skill-in-use, bound the executable consequences associated with skill-in-use, and capture evidence for attribution, verification, repair, and evolution. This paper provides a catalogue of ten empirically grounded architectural patterns (five core, five supporting) for skill harnessing and synthesises them into a reference architecture with four responsibility layers: Supply Chain, Mediation, Execution Control, and Evidence & Feedback. We evaluate the architecture through cross-instantiation across 8 selected systems. The resulting patterns and reference architecture provide a vocabulary and diagnostic frame for analysing skill-harnessing responsibilities across agent systems.

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

1 major / 1 minor

Summary. The paper claims that agent skills are static artefacts whose run-specific responsibilities (selection, binding, interpretation, evidence capture) constitute 'skill-in-use'; from analysis of eight selected LLM agent systems it derives ten empirically grounded architectural patterns (five core, five supporting), synthesises them into a four-layer reference architecture (Supply Chain, Mediation, Execution Control, Evidence & Feedback), evaluates the architecture by cross-instantiation on the same eight systems, and concludes that the patterns and architecture supply a reusable vocabulary and diagnostic frame for skill-harnessing responsibilities across agent systems.

Significance. If the patterns prove generalisable, the work supplies a structured diagnostic lens for comparing how existing and future skill-mediated agents manage the transition from static artefacts to executable, attributable runs. The empirical derivation from multiple concrete systems and the explicit layering of responsibilities are constructive contributions that could aid both analysis and design.

major comments (1)
  1. [Evaluation] Evaluation section (cross-instantiation on the eight systems): the mapping back onto the source systems demonstrates internal consistency within the sample but supplies no external test that the four layers are exhaustive or that responsibilities in architectures employing different skill-binding mechanisms, authority models, or execution semantics map without gaps or emergent patterns; this directly undercuts the claim that the architecture constitutes a reusable diagnostic frame for arbitrary LLM agent systems.
minor comments (1)
  1. [Abstract] Abstract: the phrasing 'empirically grounded architectural patterns' and 'reference architecture' could be separated more explicitly so readers immediately see which elements are observed versus synthesised.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive summary and for highlighting the significance of the work. We address the single major comment on evaluation below, agreeing that the current evaluation is limited to internal consistency and will revise accordingly.

read point-by-point responses

  1. Referee: [Evaluation] Evaluation section (cross-instantiation on the eight systems): the mapping back onto the source systems demonstrates internal consistency within the sample but supplies no external test that the four layers are exhaustive or that responsibilities in architectures employing different skill-binding mechanisms, authority models, or execution semantics map without gaps or emergent patterns; this directly undercuts the claim that the architecture constitutes a reusable diagnostic frame for arbitrary LLM agent systems.

    Authors: We agree that the cross-instantiation on the eight source systems demonstrates only internal consistency and does not provide an external test of exhaustiveness or gap-free mapping for systems with different binding mechanisms, authority models, or execution semantics. This is a genuine limitation of the current evaluation. In the revised manuscript we will (1) add an explicit 'Limitations' subsection that states the evaluation scope is confined to the sampled systems, (2) qualify the abstract and conclusion claims to describe the architecture as 'an empirically derived diagnostic frame validated on the studied systems and offered as a reusable vocabulary for further application' rather than asserting it is already proven for arbitrary systems, and (3) outline concrete directions for future external validation. These changes directly address the concern without altering the core empirical contribution. revision: yes

Circularity Check

0 steps flagged

No circularity: purely descriptive synthesis with no derivations or self-referential reductions

full rationale

The paper is a descriptive catalogue of ten patterns identified from eight selected systems, synthesized into a four-layer reference architecture and evaluated only via cross-instantiation on the same sample. No equations, fitted parameters, predictions, or derivation chains appear in the provided text. The central claim (patterns and architecture as a diagnostic vocabulary) does not reduce to any input by construction, self-definition, or self-citation load-bearing step. This is the expected outcome for architecture synthesis work that makes no mathematical or predictive claims.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no mathematical derivations, fitted parameters, or new entities described. No free parameters, axioms, or invented entities identified from available text.

pith-pipeline@v0.9.1-grok · 5734 in / 1086 out tokens · 19046 ms · 2026-06-28T22:38:14.374424+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

76 extracted references · 43 canonical work pages · 36 internal anchors

  1. [1]

    2026 , eprint =

    Skill Retrieval Augmentation for Agentic AI , author =. 2026 , eprint =

    2026

  2. [2]

    SkillClaw: Let Skills Evolve Collectively with Agentic Evolver

    SkillClaw: Let Skills Evolve Collectively with Agentic Evolver , author=. arXiv preprint arXiv:2604.08377 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization

    Skill0: In-context agentic reinforcement learning for skill internalization , author=. arXiv preprint arXiv:2604.02268 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2603.04448 , year=

    Skillnet: Create, evaluate, and connect ai skills , author=. arXiv preprint arXiv:2603.04448 , year=

    work page arXiv

  5. [5]

    Reinforcement Learning for Self-Improving Agent with Skill Library

    Reinforcement learning for self-improving agent with skill library , author=. arXiv preprint arXiv:2512.17102 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE) , pages=

    An empirical study on software bill of materials: Where we stand and the road ahead , author=. 2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE) , pages=. 2023 , organization=

    2023

  7. [7]

    Empirically-grounded reference architectures: a proposal , author=. Proceedings of the joint ACM SIGSOFT conference--QoSA and ACM SIGSOFT symposium--ISARCS on Quality of software architectures--QoSA and architecting critical systems--ISARCS , pages=

  8. [8]

    Information and software technology , volume=

    Guidelines for including grey literature and conducting multivocal literature reviews in software engineering , author=. Information and software technology , volume=. 2019 , publisher=

    2019

  9. [9]

    2007 , publisher=

    Guidelines for performing systematic literature reviews in software engineering , author=. 2007 , publisher=

    2007

  10. [10]

    Thinking with Reasoning Skills: Fewer Tokens, More Accuracy

    Thinking with Reasoning Skills: Fewer Tokens, More Accuracy , author=. arXiv preprint arXiv:2604.21764 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses

    Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses , author=. arXiv preprint arXiv:2604.25850 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering

    Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering , author=. arXiv preprint arXiv:2604.08224 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Toward Scalable Terminal Task Synthesis via Skill Graphs

    Toward Scalable Terminal Task Synthesis via Skill Graphs , author=. arXiv preprint arXiv:2604.25727 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    2026 , month = mar, day =

    The Anatomy of an Agent Harness , author =. 2026 , month = mar, day =

    2026

  15. [15]

    2026 , month = feb, day =

    Harness Engineering: Leveraging Codex in an Agent-First World , author =. 2026 , month = feb, day =

    2026

  16. [16]

    2026 , month = mar, day =

    Harness Design for Long-Running Application Development , author =. 2026 , month = mar, day =

    2026

  17. [17]

    Agent Skills Overview , author =

  18. [18]

    "Do Not Mention This to the User": Detecting and Understanding Malicious Agent Skills in the Wild

    Malicious agent skills in the wild: A large-scale security empirical study , author=. arXiv preprint arXiv:2602.06547 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    21st USENIX Security Symposium (USENIX Security 12) , pages=

    An evaluation of the google chrome extension security architecture , author=. 21st USENIX Security Symposium (USENIX Security 12) , pages=

  20. [20]

    2003 , institution=

    Capability myths demolished , author=. 2003 , institution=

    2003

  21. [21]

    ACM Transactions on Software Engineering and Methodology , volume=

    Research directions in software supply chain security , author=. ACM Transactions on Software Engineering and Methodology , volume=. 2025 , publisher=

    2025

  22. [22]

    IEEE software , volume=

    Package management systems , author=. IEEE software , volume=. 2012 , publisher=

    2012

  23. [23]

    Architectural patterns revisited-a pattern language , author=

  24. [24]

    Systems Engineering , volume=

    The concept of reference architectures , author=. Systems Engineering , volume=. 2010 , publisher=

    2010

  25. [25]

    2025 , note =

    Equipping Agents for the Real World with Agent Skills , howpublished =. 2025 , note =

    2025

  26. [26]

    2026 , note =

    Awesome Agents , howpublished =. 2026 , note =

    2026

  27. [27]

    2009 Joint Working IEEE/IFIP Conference on Software Architecture & European Conference on Software Architecture , pages=

    A classification of software reference architectures: Analyzing their success and effectiveness , author=. 2009 Joint Working IEEE/IFIP Conference on Software Architecture & European Conference on Software Architecture , pages=. 2009 , organization=

    2009

  28. [28]

    2022 , month = nov, number =

    2022

  29. [29]

    Wu, Xiyang and Li, Zongxia and Shi, Guangyao and Duffy, Alexander and Marques, Tyler and Olson, Matthew Lyle and Zhou, Tianyi and Manocha, Dinesh , year = 2026, month = apr, number =. Co-. 2604.20987 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  30. [30]

    From Skill Text to Skill Structure: The Scheduling-Structural-Logical Representation for Agent Skills

    From Skill Text to Skill Structure: The Scheduling-Structural-Logical Representation for Agent Skills , author=. arXiv preprint arXiv:2604.24026 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  31. [31]

    Wang, Jianing and Guo, Linsen and Chen, Zhengyu and Guo, Qi and Zang, Hongyu and Shi, Wenjie and Ma, Haoxiang and Xi, Xiangyu and Li, Xiaoyu and Wang, Wei and Cai, Xunliang , year = 2026, publisher =

    2026

  32. [32]

    Skills as

    Metere, Alfredo , year = 2026, publisher =. Skills as

    2026

  33. [33]

    arXiv preprint arXiv:2601.21123 , year=

    Cua-skill: Develop skills for computer using agent , author=. arXiv preprint arXiv:2601.21123 , year=

    work page arXiv

  34. [34]

    SkCC: Portable and Secure Skill Compilation for Cross-Framework LLM Agents

    SkCC: Portable and Secure Skill Compilation for Cross-Framework LLM Agents , author=. arXiv preprint arXiv:2605.03353 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    ContractSkill: Repairable Contract-Based Skills for Multimodal Web Agents

    ContractSkill: Repairable Contract-Based Skills for Multimodal Web Agents , author=. arXiv preprint arXiv:2603.20340 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2504.06821 , year=

    Inducing programmatic skills for agentic tasks , author=. arXiv preprint arXiv:2504.06821 , year=

    work page arXiv

  37. [37]

    FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills , author=

  38. [38]

    Design Principles, and Architectural Patterns for Scalable Verification (January 07, 2026) , year=

    Verifiability-First AI Engineering in the Era of AIware: A Conceptual Framework, Design Principles, and Architectural Patterns for Scalable Verification , author=. Design Principles, and Architectural Patterns for Scalable Verification (January 07, 2026) , year=

    2026

  39. [39]

    Sealing the Audit-Runtime Gap for LLM Skills

    Sealing the Audit-Runtime Gap for LLM Skills , author=. arXiv preprint arXiv:2605.05274 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  40. [40]

    oxo-call: Documentation-grounded Skill Augmentation for Accurate Bioinformatics Command-line Generation with Large Language Models

    oxo-call: Documentation-grounded Skill Augmentation for Accurate Bioinformatics Command-line Generation with Large Language Models , author=. arXiv preprint arXiv:2604.12387 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2512.23760 , year=

    Audited Skill-Graph Self-Improvement for Agentic LLMs via Verifiable Rewards, Experience Synthesis, and Continual Memory , author=. arXiv preprint arXiv:2512.23760 , year=

    work page arXiv

  42. [42]

    SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills

    Skillweaver: Web agents can self-improve by discovering and honing skills , author=. arXiv preprint arXiv:2504.07079 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  43. [43]

    Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic Tasks

    Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic Tasks , author=. arXiv preprint arXiv:2605.01293 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  44. [44]

    SkillOS: Learning Skill Curation for Self-Evolving Agents

    SkillOS: Learning Skill Curation for Self-Evolving Agents , author=. arXiv preprint arXiv:2605.06614 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  45. [45]

    MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents

    MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents , author=. arXiv preprint arXiv:2602.02474 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  46. [46]

    SkillForge: Forging Domain-Specific, Self-Evolving Agent Skills in Cloud Technical Support

    SkillForge: Forging Domain-Specific, Self-Evolving Agent Skills in Cloud Technical Support , author=. arXiv preprint arXiv:2604.08618 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  47. [47]

    Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning

    Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning , author=. arXiv preprint arXiv:2605.06130 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills

    Trace2skill: Distill trajectory-local lessons into transferable agent skills , author=. arXiv preprint arXiv:2603.25158 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2603.12056 , year=

    Xskill: Continual learning from experience and skills in multimodal agents , author=. arXiv preprint arXiv:2603.12056 , year=

    work page arXiv

  50. [50]

    GraSP: Graph-Structured Skill Compositions for LLM Agents

    GraSP: Graph-Structured Skill Compositions for LLM Agents , author=. arXiv preprint arXiv:2604.17870 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  51. [51]

    2026 , month = may, day =

    Designing, Refining, and Maintaining Agent Skills at Perplexity , author =. 2026 , month = may, day =

    2026

  52. [52]

    Workshop on Multi-Agent Learning and Its Opportunities in the Era of Generative AI , year=

    SkillTracer: Structural Failure Attribution and Refinement of Agentic Skills in Long-Horizon Web Tasks , author=. Workshop on Multi-Agent Learning and Its Opportunities in the Era of Generative AI , year=

  53. [53]

    Xiang, Keyi and Tang, Tianyi and Shao, Jie-Jing and Lyu, Yueming and Tsang, Ivor and Ong, Yew-Soon and Yin, Haiyan , booktitle =

  54. [54]

    Empirical software engineering , volume=

    Guidelines for conducting and reporting case study research in software engineering , author=. Empirical software engineering , volume=. 2009 , publisher=

    2009

  55. [55]

    Uncertainty Propagation in LLM-Based Systems

    Uncertainty Propagation in LLM-Based Systems , author=. arXiv preprint arXiv:2604.23505 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  56. [56]

    2024 , month = dec, url =

    Building Effective Agents , author =. 2024 , month = dec, url =

    2024

  57. [57]

    2026 , howpublished =

    Skill Authoring Best Practices , author =. 2026 , howpublished =

    2026

  58. [58]

    Progent: Securing AI Agents with Privilege Control

    Progent: Programmable privilege control for llm agents , author=. arXiv preprint arXiv:2504.11703 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  59. [59]

    Advances in Neural Information Processing Systems , volume=

    Swe-agent: Agent-computer interfaces enable automated software engineering , author=. Advances in Neural Information Processing Systems , volume=

  60. [60]

    International Conference on Learning Representations , volume=

    Openhands: An open platform for ai software developers as generalist agents , author=. International Conference on Learning Representations , volume=

  61. [61]

    arXiv preprint arXiv:2411.05285 , year=

    Agentops: Enabling observability of llm agents , author=. arXiv preprint arXiv:2411.05285 , year=

    work page arXiv

  62. [62]

    Understanding Automated Program Repair Agents Through the Lens of Traceability: An Empirical Study

    Understanding Software Engineering Agents Through the Lens of Traceability: An Empirical Study , author=. arXiv preprint arXiv:2506.08311 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  63. [63]

    Natural-Language Agent Harnesses

    Natural-language agent harnesses , author=. arXiv preprint arXiv:2603.25723 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  64. [64]

    arXiv preprint arXiv:2603.02176 , year=

    Organizing, orchestrating, and benchmarking agent skills at ecosystem scale , author=. arXiv preprint arXiv:2603.02176 , year=

    work page arXiv

  65. [65]

    Guanzhi Wang and Yuqi Xie and Yunfan Jiang and Ajay Mandlekar and Chaowei Xiao and Yuke Zhu and Linxi Fan and Anima Anandkumar , title =. Trans. Mach. Learn. Res. , volume =. 2024 , url =

    2024

  66. [66]

    Agent Workflow Memory , booktitle =

    Zora Zhiruo Wang and Jiayuan Mao and Daniel Fried and Graham Neubig , editor =. Agent Workflow Memory , booktitle =. 2025 , url =

    2025

  67. [67]

    CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification

    Coevoskills: Self-evolving agent skills via co-evolutionary verification , author=. arXiv preprint arXiv:2604.01687 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  68. [68]

    EvoSkill: Automated Skill Discovery for Multi-Agent Systems

    Evoskill: Automated skill discovery for multi-agent systems , author=. arXiv preprint arXiv:2603.02766 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  69. [69]

    SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution

    SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution , author=. arXiv preprint arXiv:2605.18401 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  70. [70]

    Harnessing LLM Agents with Skill Programs

    Harnessing LLM Agents with Skill Programs , author=. arXiv preprint arXiv:2605.17734 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  71. [71]

    A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications

    A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications , author=. arXiv preprint arXiv:2605.07358 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  72. [72]

    SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

    SkillsBench: Benchmarking how well agent skills work across diverse tasks , author=. arXiv preprint arXiv:2602.12670 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  73. [73]

    How Well Do Agentic Skills Work in the Wild: Benchmarking LLM Skill Usage in Realistic Settings

    How well do agentic skills work in the wild: Benchmarking llm skill usage in realistic settings , author=. arXiv preprint arXiv:2604.04323 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  74. [74]

    Counterfactual Trace Auditing of LLM Agent Skills

    Counterfactual Trace Auditing of LLM Agent Skills , author=. arXiv preprint arXiv:2605.11946 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  75. [75]

    Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks

    Skill-inject: Measuring agent vulnerability to skill file attacks , author=. arXiv preprint arXiv:2602.20156 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  76. [76]

    How Your Credentials Are Leaked by LLM Agent Skills: An Empirical Study

    Credential leakage in llm agent skills: A large-scale empirical study , author=. arXiv preprint arXiv:2604.03070 , year=

    work page internal anchor Pith review Pith/arXiv arXiv