This survey, led by researchers at Harbin Institute of Technology, reviews Agentic Tool Use from the point at which a single correct call stops being the interesting question. Early tool-learning work, including ReAct, taught models to recognize an intent and format an API request; the authors argue that practical agents now have to coordinate many tools over long trajectories, preserve intermediate state, recover from failures and operate under limits on latency, cost and safety, so that the core difficulty has become orchestration rather than access. They frame this as a qualitative change in the decision space — from choosing a tool to a series of coupled decisions covering tool-subset selection, cross-tool dependencies, sequential and parallel scheduling, failure recovery and re-planning.
The survey is motivated by two gaps the authors see in the literature. One is conceptual: terms such as tool use, tool calling, tool retrieval, workflow execution and orchestration are used loosely even though they denote different capability levels. The other is structural: planning, training, safety, efficiency, benchmarking and open-environment adaptation are usually studied separately, while deployed agent systems depend on how they interact. It therefore first gives a unified formulation of multi-tool orchestration — a policy choosing tool calls or termination from the interaction history and memory, under a cost-aware objective that trades task success against calls, latency, fees and risk — and then organizes the field around six connected dimensions.
Key Points
- The six dimensions are inference-time planning and execution, training and trajectory construction, safety and control, efficiency under resource constraints, capability completeness in open environments, and benchmark design and evaluation.
- On inference, the authors describe a shift from linear, ReAct-style sequential traces to topology-aware planning over graphs and hierarchical decompositions, and from “generate once and execute” to “search, verify, and then commit”; long-horizon orchestration additionally relies on division of labour between components, routing across large tool ecosystems, and memory that compresses or externalizes execution history.
- Tuning methods are arranged in ascending order of compute and data dependency — training-free methods (tool retrieval, prompting, non-parametric memory), multi-tool trajectory synthesis, supervised fine-tuning and reinforcement learning — and the authors argue that for synthetic data the critical measure is now coverage of long-tail and multi-step error patterns rather than quantity.
- They split multi-tool safety risks into two dimensions: parallel execution, where concurrent write operations risk race conditions and inconsistent state, and long-horizon tool chains, where malicious content or small upstream errors propagate and amplify across steps. Defences are described as moving from pre-execution constraints through transaction-style isolation and rollback to post-execution verification.
- Efficiency work is grouped into reducing latency (parallel execution, asynchronous decoupling of planning from acting, speculative execution) and reducing cost (dynamic tool retrieval instead of placing every tool description in the prompt, adaptive model routing, caching and memory).
- Capability completeness is described as a loop of perceiving capability boundaries (clarifying or terminating instead of hallucinating tools), autonomous Tool creation, and open-environment adaptation through reusable experience and skill memory.
- Benchmarks are traced from single-point call correctness to four dimensions of long-horizon capability — topological complexity of tool dependencies, temporal scale, dynamic environments, and state persistence with self-correction — and the authors argue that long-horizon evaluation should not rely solely on endpoint success.
- For applications, the authors argue that differences between domains stem mainly from their operational bottlenecks rather than from orchestration topology — for software engineering, success grounded in deterministic tests and the cost of test verification; for enterprise workflows, transactional consistency, auditing and permission boundaries.
Notes
The survey positions itself against earlier surveys of tool learning and LLM agents by taking multi-tool orchestration, rather than tool use in general or agent systems broadly, as its unit of analysis, and by drawing boundaries between concepts it considers often conflated — tool invocation, tool retrieval, orchestration and toolset expansion. Its closing agenda asks for better abstractions for stateful orchestration, stronger evaluation protocols for dynamic and long-horizon settings, and closer integration of model-level reasoning with system-level guarantees; for applications it lists adaptive regression against interface drift, process-level verifiability and replayable audits, joint optimization of cost, latency and quality, and composable, adversarially robust failure recovery.
The version extracted here is arXiv:2603.22862v2 [cs.SE], formatted with an ACM template whose venue fields are still placeholders and a stated publication date of April 2026. The authors are affiliated with Harbin Institute of Technology, Harvard University and Huawei Technologies.