action planning

In the era of large language models, agent workflows face critical challenges in scalability, controllability, and security. To address these, this paper presents a systematic literature review and proposes, for the first time, a dual-dimensional taxonomy—spanning functional capabilities (task planning, multi-agent collaboration, tool integration) and architectural characteristics (role definition, orchestration process, specification languages). Through comparative analysis of over twenty representative academic and industrial systems, we identify recurring design patterns and persistent technical bottlenecks. We further introduce security-enhanced orchestration optimization strategies and pinpoint core gaps, including the lack of standardization and insufficient multimodal integration. This work establishes a foundational theoretical framework and practical guidelines for the design, evaluation, and evolution of agent workflows, advancing the field toward structured, trustworthy, and multimodal-cooperative paradigms.

Existing BDI agent planning approaches rely heavily on manual design and are predominantly limited to single-agent settings, rendering them inadequate for modeling collaborative or competitive dynamics in multi-agent systems. Method: This paper introduces Alternating-Time Temporal Logic (ATL) into the BDI framework for the first time, proposing an automated plan generation method for multi-agent systems. Group-level objectives—such as “cooperatively achieve” or “competitively prevent”—are formally specified in ATL, and model checking is employed to synthesize BDI plans satisfying inter-agent interaction constraints. Contribution/Results: The approach significantly reduces human intervention and supports dynamic, asymmetric multi-agent interactions. Empirical evaluation on cooperative game tasks demonstrates that the generated plans effectively coordinate agents toward shared goals, validating the method’s feasibility and practical utility.

This work addresses the long-standing challenge of efficiently compiling temporal numeric planning problems with continuous actions into PDDL+. Existing approaches struggle to achieve this transformation while preserving semantics and maintaining tractable model size. The paper proposes a practical polynomial-time compilation method that, under the mild assumption that actions are non-self-overlapping, exactly translates such problems into discrete PDDL+ models. The resulting encoding fully preserves the original semantics and incurs only a constant-factor increase in plan length. To the best of our knowledge, this is the first compilation technique that simultaneously guarantees semantic fidelity, polynomial scalability, and applicability to general temporal numeric planning, thereby filling a significant gap in the literature. Empirical evaluation demonstrates that the proposed compilation substantially enhances the feasibility of solving complex temporal numeric planning problems using existing PDDL+ planners.

This work addresses the limited cross-site generalization of existing large language model–based web agents, which stems from their flat memory architectures that tightly couple task logic with site-specific operations. To overcome this, the authors propose the Hierarchical Memory Tree (HMT) framework, which explicitly decouples logical planning from action execution for the first time. HMT constructs a three-tiered structure through automated trajectory abstraction: an intent layer unifying task objectives, a phase layer defining reusable semantic subgoals with pre- and post-condition constraints, and an action layer storing transferable element-level semantic descriptions. Integrated with phase-aware reasoning and condition validation mechanisms, HMT significantly outperforms current methods on the Mind2Web and WebArena benchmarks, demonstrating markedly enhanced robustness and generalization—particularly in cross-site and cross-domain scenarios.

To address the challenge of balancing concurrent action conflicts and plan executability in multi-agent task planning, this paper proposes a two-stage LLM-PDDL collaborative framework. First, a large language model (LLM) performs commonsense-driven goal decomposition to generate mutually exclusive, parallelizable sub-goals. Second, a classical PDDL planner (e.g., FF or Fast Downward) independently synthesizes formally verifiable single-agent plans for each agent. This work is the first to integrate the LLM’s high-level goal abstraction capability with the formal correctness guarantees of classical planning. Empirical results demonstrate 100% action executability, significantly reduced planning time, and plan step counts that outperform single-agent baselines while approaching human expert performance—thereby unifying efficiency, feasibility, and coordination quality in multi-agent planning.

To address the challenge of executing multimodal user instructions that require coordinated invocation of heterogeneous models under diverse constraints, this paper proposes HIVE: a framework grounded in LLM-driven PDDL-based formal planning. HIVE decomposes complex queries into verifiable atomic action sequences and dynamically orchestrates specialized multimodal models accordingly. Its methodology integrates knowledge-aware task decomposition, constraint-guided action planning, and model orchestration to ensure end-to-end interpretability and traceability. Evaluated on the MuSE benchmark, HIVE establishes new state-of-the-art performance in both task selection and multimodal planning, significantly outperforming existing collaborative systems. Crucially, HIVE introduces the first LLM+PDDL joint planning paradigm for multimodal, cross-model tasks—enabling strong constraint satisfaction and formal verification of execution plans.

This work addresses the limitations of existing large language model (LLM) agents in complex, long-horizon web tasks, where constrained context memory, inadequate planning capabilities, and greedy action selection often lead to premature failure. To overcome these challenges, we propose STRUCTUREDAGENT, a novel framework that integrates a dynamic AND/OR tree with a structured memory mechanism to enable interpretable, online hierarchical planning. The AND/OR tree facilitates efficient search and maintenance of candidate solutions, while the structured memory enhances long-term reasoning and fault tolerance. Extensive experiments on WebVoyager, WebArena, and a custom shopping benchmark demonstrate that our approach significantly outperforms standard LLM agents, achieving substantial improvements in both constraint satisfaction and overall task success rates.

Existing multi-agent systems rely heavily on manual design, resulting in low efficiency and poor scalability. This work proposes an end-to-end automated framework that leverages a large language model (LLM) as a planner to decompose tasks, coupled with a two-stage retrieval-based agent recommendation mechanism—integrating embedding similarity and reranking—to achieve precise task-to-agent matching. To ensure workflow consistency and robustness, the framework incorporates a dynamic invocation graph and a global critical evaluator agent. Experimental results demonstrate that the proposed approach significantly improves task completion recall and outperforms existing methods in both scalability and overall performance, with the critical evaluator further enhancing plan quality.

This work addresses the challenge of interpreting and analyzing long-horizon autonomous agent behaviors, which are typically recorded as unstructured natural language traces that hinder efficient diagnosis of inefficiencies, errors, and compliance violations. To overcome this limitation, the authors introduce ACT*ONOMY—the first grounded theory–based behavioral taxonomy for autonomous agents—featuring a three-level hierarchical structure comprising 10 actions, 46 sub-actions, and 120 leaf categories. Accompanied by an open knowledge base and an automated annotation pipeline, ACT*ONOMY provides a scalable, shared semantic vocabulary and standardized analysis protocol. This framework enables systematic cross-agent behavior comparison, facilitates failure mode identification, and reveals behavior patterns correlated with system faults, thereby substantially enhancing human oversight and control over autonomous agents.

This work addresses the challenges of weak semantic coherence, incomplete actions, and temporal drift in long-horizon video generation with open-loop models. To overcome these limitations, the authors propose ActWM, a closed-loop, self-improving action world model that implements a “plan–execute–reflect” framework. Specifically, a PlanAgent decomposes high-level semantic actions into object-centric sub-actions, while a CriticAgent provides iterative evaluation feedback enriched with long-term memory. The system is further refined through reinforcement learning to enable continuous optimization. This approach significantly enhances semantic alignment and temporal consistency, delivering consistent performance gains across multiple text-to-video backbone architectures. The effectiveness of ActWM is rigorously validated on a newly curated benchmark, comprising the ActWM-Dataset and ActWM-Bench.