2026· Annual Meeting of the Association for Computational Linguistics· pp. 24960-24969· 1 citation· 32 references
Computer Science
TL;DR
HFlow is proposed, an evolutionary framework for generating agentic workflows through human-agent collaboration that produces cost-effective and adaptive solutions, better aligning automated agen-tic systems with dynamic human needs.
Abstract
Agentic workflows, composed of multiple collaborating Large Language Models (LLMs), have become a key paradigm for complex problem-solving. However, their effectiveness is often hindered by three critical challenges: high manual design costs, inefficient agentic search, and poor dynamic adaptability to new tasks and human preferences. To address these limitations, we propose HFlow, an evolutionary framework for generating agentic workflows through human-agent collaboration. HFlow employs an evolutionary algorithm to automate the search for optimal workflows by mutating and crossing over their structures, prompts, and LLM backbones. This process is guided by human preferences to ensure rapid convergence, while a hierarchical experience memory enables the generalization of learned strategies. Extensive experiments on math and code generation benchmarks show HFlow surpasses other automated baselines by up to 27.34%, while achieving comparable performance to o1-preview at only one-fourth of the cost. Our work introduces a new paradigm for workflow design that produces cost-effective and adaptive solutions, better aligning automated agen-tic systems with dynamic human needs.
This work introduces AgentSearch, a cost-aware Monte Carlo Tree Search (MCTS) framework that constructs agentic workflows through deliberative lookahead search and attains single-episode success while reducing computational costs by up to 47%, thereby eliminating the trial-and-error exploration required by previous adaptive methods.
D. Attota, Ying Xie· Proceedings of the 15th Inte...· 0 citations
ReASearch is presented, a unified framework for reasoning-driven optimization in which the agent autonomously decides what to evaluate, how to diagnose failures, which edits to make, and when to verify or restart.
EASy is proposed, a trainable agentic framework that jointly optimizes task performance and computational efficiency through reinforcement learning and consistently achieves stronger performance-efficiency trade-offs than strong agentic baselines.
Junnan Liu, Linhao Luo, Thuy-Trang Vu et al.· 0 citations
For our Special Interest Group (S.I.G.), we propose papers that address the orchestration of teams by synthesizing their workflows into a coherent whole, whether these teams are composed of human, machine, Generative AI (gen-AI), robot or AI-Agentic members. The bigger picture of interdependence, teamwork and Gen-AI indicates the need by organizations to build a library of human and artificial agents with bidirectional agency (responsibility) to achieve operational goals (missions), considering agentic risk tolerances, available skills, and vulnerabilities across a complex trade space among the skills available versus those needed for the tasks assigned to complete an operation. In this trade space, agents (human or artificial) from multiple systems with the requisite skills to accomplish a designated task and timeline combined to form a hierarchy of humans, robots, machines and AI. This complex system produces workflows that must be synthesized into a unit(s), then orchestrated to accomplish the goals assigned to it, yet remain trusted even in competitive and uncertain environments. Once synthesized into a unit (e.g., a team), Gen-AI provides the opportunity to not only advance the science of teams by orchestrating team products and performances, but also has raised several concerns (viz., AI used for deception, superintelligence, blackmail, or existential threats to humans). For our S.I.G., We are interested in orchestrating teams: What are the benefits, drawbacks, and, most importantly, can humans, machines and Agentic AI be synthesized and managed (orchestrated)?
W. Lawless, Marco Brambilla, Stephen Russell· AHFE International· 0 citations
Recent advances in large language models (LLMs) have enabled the development of increasingly sophisticated agentic AI systems capable of reasoning, planning, tool use, memory management, and collaborative problem solving. Agentic AI refers to AI systems that can autonomously pursue goals and execute tasks by combining reasoning, decision-making, memory, interaction with external tools, and adaptive workflows. Unlike traditional LLM applications that primarily rely on single-prompt interactions or fixed sequential pipelines, agentic AI systems can iteratively analyze problems, coordinate multiple steps, invoke external resources, refine intermediate outputs, and dynamically adapt their behavior in response to evolving context and feedback. Recent developments in agent frameworks and orchestration platforms have accelerated interest in building such systems for applications involving analytics, research assistance, automation, and decision support. Despite rapid advances in LLM capabilities, designing robust and reliable agentic workflows remains challenging. Developers must address issues related to state management, workflow orchestration, tool integration, multi-agent coordination, reasoning reliability, and transparency of decision-making processes. Recent frameworks such as LangGraph offer new opportunities to build stateful, graph-based AI systems that support complex reasoning workflows, adaptive execution paths, and collaborative multi-agent architectures [1]. This hands-on tutorial introduces LangGraph, a framework built on top of LangChain for designing and orchestrating stateful agentic AI workflows [2]. LangGraph enables developers to construct graph-based systems in which nodes represent reasoning steps, tools, memory modules, or specialized agents, while edges define execution flow and state transitions. Compared with traditional linear LLM pipelines, graph-based architecture offers greater flexibility in implementing branching logic, iterative refinement, recovery mechanisms, memory-aware execution, and collaborative multi-agent interactions. The tutorial focuses on practical techniques for designing and implementing modern agentic AI systems that move beyond simple prompt engineering toward structured reasoning workflows capable of autonomous decision making and coordinated execution. Participants will learn how to build multi-step reasoning pipelines, integrate external APIs and computational tools, manage workflow state and conversational memory, and design collaborative systems involving multiple specialized agents with distinct responsibilities. The tutorial also discusses practical considerations in workflow orchestration, reliability, scalability, evaluation, and deployment of agentic AI systems in real-world environments. Through guided coding exercises implemented in Python using LangGraph, LangChain, and modern LLM APIs, attendees will progressively construct increasingly sophisticated systems. Hands-on activities include building tool-using agents for data analysis and information retrieval, implementing collaborative reasoning systems composed of specialized agents, developing self-refinement and reasoning loops, and constructing decision-support agents that integrate qualitative and quantitative information to generate structured recommendations. Participants will also explore methods for designing transparent and human-centered agentic workflows suitable for real-world applications. The tutorial emphasizes practical system design for data science, analytics, research assistance, and decision-support applications. In addition to implementation techniques, the tutorial examines challenges associated with hallucinations, automation bias, evaluating agent behavior, the reliability of reasoning chains, and the responsible deployment of AI systems that autonomously interact with tools and external information sources. The tutorial encourages participants to adopt human-centered AI principles and design practices that promote transparency, accountability, and responsible use of agentic systems. The tutorial combines conceptual discussions, live coding demonstrations, and interactive, hands-on exercises for data scientists, machine learning engineers, AI researchers, and software developers interested in building scalable, practical agentic AI systems. By the end of the tutorial, participants will have implemented a complete multi-agent decision assistant capable of gathering information, coordinating specialized agents, evaluating alternatives, and generating structured recommendations using modern graph-based AI workflows.
Mohammad Amin Kuhail· Proceedings of the 32nd ACM...· 0 citations
TestAgent is proposed, an LLM-based test generation approach that addresses the above limitations by emulating human testing practices via a multi-agent collaboration mechanism and equips TestAgent with a set of tool APIs that can be invoked dynamically in an on-demand and adaptive manner.
Quanjun Zhang, Ye Shang, Siqi Gu et al.· 0 citations