tool-augmented reasoning

Current large language model (LLM) agents face challenges in real-world deployment, including inefficiency, error-proneness, and poor maintainability, largely due to their reliance on on-the-fly reasoning and low-level tool invocation. This work introduces, for the first time, a skill-centric agent architecture that formalizes a comprehensive skill lifecycle framework encompassing representation, acquisition, retrieval, and evolution. It positions skills as a complementary mechanism bridging high-level reasoning and operational execution. By integrating key techniques—such as skill representation learning, automated acquisition, semantic retrieval, and continual evolution—and synergizing them with tool use, memory mechanisms, and contextual constraints, the proposed framework establishes a reusable and composable skill system. The paper further surveys representative approaches, open-source resources, and application scenarios, offering both theoretical foundations and practical guidance to enhance the scalability, robustness, and maintainability of intelligent agent systems.

This work addresses the high computational cost, latency, and failure rates in AI agent workflows caused by redundant reasoning and repetitive tool invocations, often exacerbated by hallucinations. To mitigate these issues, we propose the Agent Workflow Optimization (AWO) framework, which, for the first time, automatically identifies recurring tool-calling patterns through trajectory analysis and abstracts them into deterministic meta-tools. This abstraction reduces the number of intermediate large language model (LLM) calls, thereby enhancing both execution efficiency and robustness. Our approach enables fully automated optimization of agent workflows, achieving up to an 11.9% reduction in LLM invocations and a 4.2 percentage point improvement in task success rate on two widely used benchmarks.

This work investigates how collaborative architecture design affects collective reasoning in multi-agent large language model (LLM) systems, with a focus on expertise allocation as a critical bottleneck. Methodologically, we conduct systematic ablation studies examining the interplay among three dimensions: domain-aligned expert specialization, collaboration paradigms (structured workflows vs. diversity-driven knowledge fusion), and system scale. Our results show that domain-dependent expert alignment substantially improves reasoning accuracy; diversity-aware knowledge integration outperforms rigid task decomposition; and communication overhead constitutes the primary scalability bottleneck. Based on these findings, we propose a configurable multi-agent design framework that quantifies the compute–performance trade-off under scale expansion and empirically validates the significant gains from expert alignment on context-intensive reasoning tasks.

This paper addresses foundational challenges in advancing large language models toward autonomous agents—specifically, the formal definition of autonomy, decision-making mechanisms, and goal hierarchies. Method: It introduces the principle “knowledge boundary equals tool-use boundary,” unifying internal reasoning and external tool invocation as equivalent cognitive operations; cognitive action boundaries are rigorously defined by knowledge accessibility, enabling optimal cognitive efficiency. Integrating cognitive modeling, knowledge representation, and decision theory, the approach constructs a unified, interpretable cognitive framework tailored for foundation agents. Contribution/Results: The framework enables adaptive, goal-directed, and minimally redundant tool invocation. It establishes the first theoretically grounded, architecture-agnostic cognitive paradigm for autonomous agents—providing both rigorous formal foundations and a scalable design blueprint for next-generation intelligent systems.

This work addresses the challenge of coordinating multiple tools with large language models to accomplish long-horizon tasks in complex, dynamic environments—a setting where prior research has largely been confined to single-tool or one-step interactions. We formulate multi-tool coordination as dynamic orchestration over extended trajectories and introduce a unified task definition alongside a structured analytical framework. Through six core dimensions—planning and execution, training methodologies, safety, efficiency, capability completeness, and evaluation benchmarks—we systematically review key techniques, including reasoning-time planning, trajectory-based training, safety-aware control, resource-constrained optimization, and modeling in open-ended environments. By synthesizing insights from software engineering, enterprise workflows, GUI automation, and mobile systems, we delineate the fundamental challenges and future directions for building reliable, scalable, and verifiable multi-tool agents.

Production-grade autonomous AI workflows face significant engineering challenges in reliability, observability, maintainability, and security governance. Method: We propose a structured, full-lifecycle methodology comprising a multi-agent architecture with collaborative reasoning, tool augmentation, and dynamic orchestration—integrated with the Model Context Protocol (MCP), deterministic orchestration, pure function invocation, containerized deployment, and modular tool integration. We further define nine core engineering practices, including tool-first design, single-responsibility agents, externalized prompt management, and model-federation-driven responsible AI design. Contribution/Results: This work establishes the first systematic engineering paradigm for Agentic AI productionization, markedly improving system simplicity, observability, and governability. Empirical validation via a multimodal news analysis–media generation use case demonstrates robustness and scalability. The methodology provides a reusable framework and practical benchmark for industrial-scale autonomous AI systems.

This work addresses the lack of systematic methodologies in developing large language model (LLM) agents for scientific domains, the misalignment between domain experts and developers in understanding constraints, and the uneven capabilities arising from LLMs’ “jagged technological frontier.” To tackle these challenges, the paper proposes a tripartite, stage-gated agent engineering paradigm that integrates structured requirement templates, tool orchestration mechanisms, and multi-stage validation gates. This approach facilitates close collaboration among domain experts, developers, and auxiliary agents to transform informal intents into auditable, testable, and maintainable agent specifications. Empirical evaluation in scientific application scenarios demonstrates substantial improvements in both development efficiency and complex query performance, thereby validating the effectiveness of the proposed framework in enhancing agent specifiability, testability, and maintainability.

Existing multimodal agents are constrained by static toolkits, limiting their generalization to novel scenarios and often introducing redundancy and errors through indiscriminate tool invocation. This work proposes the first self-evolving multimodal agent framework capable of online skill forging and reuse, integrating four tightly coupled stages—decision, retrieval, adaptation, and forging—to dynamically determine whether to use tools, select existing ones, or synthesize new skills on demand, thereby establishing a closed-loop evolutionary mechanism. The framework employs a unified policy to dynamically choose among direct answering, tool reuse, or skill forging, and leverages reinforcement learning to jointly optimize the necessity of invocation, retrieval accuracy, execution effectiveness, and skill reusability, augmented with explicit cost penalties to suppress redundant calls. Evaluated across twelve benchmarks, the approach significantly outperforms sixteen baselines, achieving notable advances in accuracy, efficiency, and generalization.

This work investigates whether tool-augmented reasoning outperforms native chain-of-thought (CoT) under semantic interference and elucidates the underlying mechanisms of performance degradation. By employing a factorized intervention framework, the study disentangles the costs of prompt formatting, overhead from tool-calling protocols, and actual utility derived from tools, introducing for the first time the concept of a “tool usage tax” to quantify the performance penalty inherent to the protocol itself. To address this, the authors propose G-STEP, a lightweight gating mechanism that dynamically modulates tool invocation during reasoning. Experiments reveal that under semantic noise, the benefits of tool use often fail to offset protocol-induced overhead, resulting in performance inferior to native CoT. While G-STEP partially mitigates this issue, substantial improvement ultimately hinges on enhancing the model’s intrinsic reasoning capabilities.

Existing multi-agent collaborative systems are hindered by static workflows, sequential scheduling, and heterogeneous interfaces, leading to high complexity and poor scalability. This work proposes Agent-as-Tool, a unified paradigm that abstracts both agents and tools into a standardized, learnable action space, and introduces ParaManager—a lightweight coordinator enabling state-aware parallel subtask decomposition, delegation, and asynchronous execution. By unifying communication protocols and incorporating explicit state feedback, the framework facilitates efficient multi-agent collaboration. A two-stage training strategy—combining supervised fine-tuning with a recovery mechanism and reinforcement learning—optimizes task success rate, protocol compliance, response diversity, and reasoning efficiency. Experiments demonstrate that ParaManager achieves strong performance across multiple benchmarks and exhibits robust generalization to unseen agent pools.

This work addresses the vulnerability of existing large-scale multi-agent systems to failure in complex tasks due to error propagation and insufficient verification mechanisms. The authors propose a two-stage framework that automatically constructs and executes task-specific multi-agent systems from natural language instructions, incorporating dual verification mechanisms—during both construction and runtime. The approach decomposes tasks into directed acyclic graphs, defines input/output contracts, grounds knowledge via web search, and auto-generates prompts and tools. It further introduces a three-level error attribution scheme and intermediate output validation gating to enable targeted recovery strategies. Experimental results demonstrate that the method significantly outperforms strong baselines across programming, in-context learning, and open-ended reasoning tasks, consistently improving task success rates, error recovery capability, and workflow stability.