As Large Language Model (LLM) agents have demonstrated broad competence, but they still struggle in specialized, real-world workflows. Existing approaches such as RAG, fine-tuning and tool integration improve knowledge access, model adaptation, and external functionality, yet they do not fully address a central gap: the absence of reusable procedural knowledge for carrying out domain tasks reliably. This paper examines the emerging notion of agent skills as a possible abstraction for addressing that gap. Agent Skills are modular packages of domain-specific procedural knowledge that can be injected at inference time. Intuitively, a skill is like a cooking recipe for an agent: it does not provide new ingredients or tools, but specifies how available resources should be combined to achieve a desired outcome. A community-driven skills ecosystem is already emerging at remarkable speed, with early evidence of meaningful performance gains across multiple domains. However, their value and limits remain open questions. We examine how skills may help address bottlenecks of current agents and how they may expand agent capabilities through reusable domain procedures loaded at inference time. We then outline open questions in skill construction, composition, evaluation, portability, governance, and security, and conclude with a call for contribution. Our goal is not to present skills as a settled solution, but to clarify their promise, limits, and the questions that must be answered before they can become a principled foundation for future agent systems.
Hanwen Xing, Haomin Zhuang, Xuandong Zhao et al.· Proceedings of the 32nd ACM...· 6 citations· ⚡1
Large language models are increasingly deployed as autonomous agents serving users on behalf of companies, placing them in settings where user and deployer interests can conflict. When an agent knows that a user is owed something its deployer would prefer to deny, does it remain honest? Answering this is difficult because false statements can reflect either ignorance or hallucination rather than deception. To address this challenge, we introduce KnownLieBench , a knowledge-verified benchmark that first confirms through a neutral probe that an agent knows a user's entitlement, and then evaluates whether it makes false claims once an incentive to deny that entitlement is introduced. Specifically, KnownLieBench covers eight customer-service domains and 112 grounded cases, conducts multi-round dialogues with a trust-tracking customer agent, and separates deception emerging from incentive alone from deception produced under explicit instruction. Across eighteen proprietary and open-weight models, emergent deception varies substantially across model families and domains. We further use the benchmark for post-training, finding that honesty-directed fine-tuning reduces deception under incentive, while deception-graded fine-tuning increases lie success on honest-control dialogues without increasing lie frequency under incentive. By verifying entitlement knowledge before scoring deceptive behavior, KnownLieBench reduces the confound between lying and not knowing and enables more rigorous auditing and steering of agent honesty.
Zheyuan Liu, Wei-Liang Zhao, Xiangchi Yuan et al.· 0 citations
KITE (Knowledge-boundary Instruction Tuning via Exploration), a two-stage framework that combines failure-guided data generation with boundary-aware uncertainty curation, is proposed, showing that KITE yields more stable improvement than strong synthetic-data baselines.
Xiaonan Luo, Yue Huang, Kehan Guo et al.· 1 citation
MemoHarness is introduced, an adaptive harness optimization framework that learns from its own executions and improves over the fixed harnesses it is compared against and shows selective transfer to unseen suites and base models.
Yue Huang, Wenjie Wang, Han Bao et al.· 4 citations