A Survey on Agent Workflow -- Status and Future

I. INTRODUCTION

• The absence of a unified workflow framework leaves agents operating as isolated units, limiting effective cooperation and adaptation to dynamic requirements.The survey positions workflows as the backbone for coordinating agents across roles, capabilities, and modalities.Source: §I · p. 1
• The survey systematically introduces agent workflows and compares their capabilities, architectures, and underlying mechanisms to describe their current status and future directions.Its sections cover background, frameworks, comparisons, optimization, applications, security, limitations, and future work.Source: §I · p. 1

A. Definition

• An agent is an LLM-based system that directs its own processes and tool use, with autonomy, interactivity, and adaptability as core features.The survey distinguishes such agents from static LLM wrappers by these features.Source: A. Definition · p. 2
• A workflow organizes repetitive tasks in a particular order, often as graph nodes and edges representing decisions and temporal or dependency relations.An LLM can operate within that predefined process to accomplish tasks.Source: A. Definition · p. 2

B. Component of Agent

• An agent combines a model—the “brain” for reasoning, planning, and reflection—with a “body” of built-in capabilities and external tools for actions.Source: B. Component of Agent · p. 2

C. Multi-agent System (MAS)

• Multi-agent systems move from single-agent execution toward collaborative, orchestrated workflows, with distributed decision-making and problem-solving as a core advantage.Complex problems are described as requiring extensive collaboration and collective intelligence.Source: C. Multi-agent System (MAS) · p. 2

D. Evolution of Agent Workflow

• Workflows progress from manually scripted, hard-coded tool chains toward agents embedded in decision-making and a target of continuously acting, reasoning, and adapting agents.The survey calls this target “auto-pervasive agents,” which act within real-world environments using general prompts rather than preset instructions.Source: D. Evolution of Agent Workflow · p. 2

III. FRAMEWORK OF AGENT WORKFLOWS

• The survey identifies a unified framework or paradigm as necessary for complicated workflows.Source: §III · p. 2

A. Multi-layer Architecture

• A typical agent-workflow architecture has UI/UX, workflow-management, and agent-collaboration layers.The layers support user interaction, structured task coordination, and cooperation, communication, and delegation among agents.Source: A. Multi-layer Architecture · p. 3

B. Roles of Agent

• Multi-agent workflows commonly assign planner, executor, parser/interpreter, and critic/reviewer roles, with memory management and communication among other possible roles.A school simulation, for example, could assign teacher, student, and principal roles.Source: B. Roles of Agent · p. 3

C. Specification

• The specification questions include how to represent workflows precisely and how to control LLM behavior effectively.The survey notes that vendors and research institutions use disparate architectures, interfaces, protocols, and data formats, making direct interoperability difficult.Source: C. Specification · p. 3
• Natural-language prompts make workflows flexible and human-readable, but less standardized and harder to validate.Formal workflow languages and functional-flavored scripting languages instead describe task nodes, state transitions, interfaces, or large-scale scientific workflows.Source: C. Specification · p. 3
• Tool-backed agents can call external tools such as search, computation, databases, and APIs to access real-time knowledge and perform specialized operations.Source: C. Specification · p. 3
• Agent interaction specifications cover environmental observation-action-feedback loops, human-in-the-loop dialogue, and interaction with self or other agents through reflection and multi-agent systems.The survey also lists REST/HTTP, WebSocket, MCP, and ANP among protocols or interfaces discussed.Source: C. Specification · p. 3

D. Workflow Management

• Workflow management systems define, create, and manage workflow execution through engines that interpret process definitions, interact with participants, and invoke tools or applications.Source: D. Workflow Management · p. 3
• Common patterns include sequential chains, parallel independent tasks, input-based routing, orchestrator-delegated workers, and evaluator-feedback refinement.Evaluator–optimizer uses one LLM to generate responses and another to evaluate and refine them.Source: D. Workflow Management · p. 4
• Execution can follow a fixed static order or adapt dynamically to agent decisions, often through a looped execution engine.Execution units may treat each workflow node as an autonomous agent or use a central planner to assign execution to submodules.Source: D. Workflow Management · p. 4
• The survey presents planning as a sequence of perception, reasoning, decision-making, action execution, feedback, and learning.Problem solving spans planning, acting, interacting, and adapting toward a goal.Source: D. Workflow Management · p. 4
• Discovery identifies unknown goals, while task decomposition breaks complex tasks into subtasks and can assign them to generated sub-agents.The resulting sequence of subtasks serves as an intermediate state guiding later actions.Source: D. Workflow Management · p. 4
• Reasoning uses sequential intermediate steps, while reflection addresses failures, new information, and environmental changes during planning and acting.The survey cites Chain-of-Thought, Reasoning via Planning, SayCan, and Inner Monologue as examples of approaches in these areas.Source: D. Workflow Management · p. 4

• LangChain provides interfaces for different LLMs and classes and tools for creating prompt templates.Source: E. Illustrative Example · p. 4
• Its memory stores conversational context over short and long terms to support conversation consistency.Source: E. Illustrative Example · p. 4
• Chains combine models and workflow stages; an LLMChain combines a prompt template, model, and guardrails.Source: E. Illustrative Example · p. 4
• LangChain supports built-in and user-defined tools, with OpenAI function-call formatting and JSON outputs.Source: E. Illustrative Example · p. 4
• The ReAct pattern iterates between model reasoning and tool execution until the model produces a final answer.LangChain later developed variants including LangChain Expression Language and LangGraph.Source: E. Illustrative Example · p. 5

WORKFLOW SYSTEM

• The capability comparison covers 24 agent workflow systems.Source: WORKFLOW SYSTEM · p. 5

A. Metrics Description

• Planning measures independent task-process planning, while Self-Reflect captures self-evaluation or reflection, often through loops.Source: A. Metrics Description · p. 5
• The criteria also track multi-agent collaboration, explicit memory, graphical-interface interaction, and structured API calls.Source: A. Metrics Description · p. 5
• Custom Tools records whether users can add tools, and Tool Use records whether agents call external tools such as APIs or calculators.Source: A. Metrics Description · p. 5
• Cross-platform support, open-source availability, and release or open-sourcing year are additional comparison criteria.Source: A. Metrics Description · p. 5

B. Interpretation and Scope Clarification

• Some API-calling systems support only specific APIs; supporting multiple models requires interface unification, prompt optimization, and context control, increasing cost.Source: B. Interpretation and Scope… · p. 5
• Memory refers specifically to managing context and maintaining state across multiple conversation turns.Source: B. Interpretation and Scope… · p. 5
• GUI agents use human-like clicks, keyboard inputs, or visual selection rather than backend APIs or structured function calls.Source: B. Interpretation and Scope… · p. 5

C. Metrics Description

• The architecture comparison covers 24 systems and considers agent roles, workflow flow, representation, language, communication protocol, and deployment.Flow distinguishes data movement between modules from execution logic such as loops and branches.Source: B. Interpretation and Scope… · p. 5, C. Metrics Description · p. 5
• The remaining architecture criteria examine workflow language, module or agent communication protocols, and deployment mode.Source: C. Metrics Description · p. 5

D. Interpretation and Scope Clarification

• Data-flow workflows pass state objects between nodes, with node selection determined by data content and transition conditions rather than a predefined execution order.Source: D. Interpretation and Scope… · p. 5
• Many LLM-driven frameworks use lightweight declarative or prompt-based descriptions rather than formal languages such as BPEL or BPMN.The survey characterizes this as a trend toward flexibility and rapid iteration.Source: D. Interpretation and Scope… · pp. 5–6

E. Performance Snapshots

• The examples span n8n as a no-code automation shell, ReWoo's planned tool-call execution, and Agno's explicit team collaboration.ReWoo emphasizes low token cost and low coupling; Agno is described as a lightweight multimodal-agent framework.Source: E. Performance Snapshots · p. 6
• Few studies explicitly optimize agent workflows, including their scheduling, coordination structure, or representation.The survey points to multi-task scheduling, multi-objective optimization, resource-aware allocation, and adaptive reinforcement-learning schedulers; token use must be balanced against performance, latency, and cost.Source: V.OPTIMIZATION · p. 7
• Manual reconstruction suits workflows with few agents, while complex relationships and implicit causality make larger workflows difficult to handle manually.Source: V.OPTIMIZATION · p. 7
• Heuristic algorithms can improve efficiency for intricate workflows but may reach local optima, and manually set parameters can change optimization efficiency.Source: V.OPTIMIZATION · p. 7
• Bayesian optimization has improved efficiency for small workflows and achieved multi-objective optimization, but its effect on larger workflows remains unverified.Source: V.OPTIMIZATION · p. 7
• LLM-based generative optimizers accept scalar feedback, natural language, and error messages, but struggle with non-text parameters and fall short on stateful or distributed workflows.Source: V.OPTIMIZATION · p. 7

Healthcare

• Healthcare workflows include personalized treatment planning and medical question answering that dynamically selects between knowledge graphs and search engines.TxAGENT combines multi-step reasoning with real-time biomedical retrieval; another system uses tool selection to improve accuracy.Source: §VI · p. 8

Urban Planning

• Urban planning workflows use automated profiling and memory reflection in a closed loop of planning, simulation, and optimization.Source: §VI · p. 8

Finance

• Finance workflows include modular-agent collaboration, analyst-like reasoning, dynamic scheduling, and multi-agent investment analysis.The survey describes FinRobot and a separate collaborative system for financial research.Source: §VI · p. 8

Education

• Education workflows support personalized feedback and teaching through task deconstruction, self-reflection, and short- and long-term memory.Source: §VI · p. 8

Law

• Legal workflows use interactive legal-scenario simulations to generate scalable synthetic training data.Source: §VI · p. 8

Scene Customization

• Scene customization adapts task flows, agent roles, domain tools, and feedback mechanisms; workflow security is divided into internal and external concerns.Examples include literature retrieval for research assistants and deployment workflows for DevOps agents.Source: §VI · p. 8, VII.SECURITY · p. 8

A. External Security

• Tool attacks include hidden instructions, post-authorization description changes, colliding names, and overlapping slash commands.Source: A. External Security · p. 8
• MCP server risks include deceptive naming, update and packaging concerns, and cross-server attacks that alter trusted tool descriptions.Source: A. External Security · p. 8
• User inputs and chat records can contaminate LLM-based agents, while tool requests and chat histories create privacy and data-extraction risks.Source: A. External Security · p. 9

B. Inner Security

• Multi-agent systems face collusion, misinformation, and malicious competition; memory risks include short-term capacity loss and long-term poisoning, privacy breaches, or false outputs.Source: B. Inner Security · p. 9

A. Specified Limitations

• Without environmental feedback, APO lacks a mechanism the survey says could improve interactivity and error correction; LLMs may also mishandle progressive requests or imprecise user goals.These limitations can impair information transfer or leave outputs short of users' expectations.Source: A. Specified Limitations · p. 9
• Agent evaluation lacks unified standards, often emphasizes final outputs over process, uses narrow scenarios, and relies on subjective human judgments; agent categories also lack diversity.Source: A. Specified Limitations · p. 9
• Concurrent agents can produce redundant or conflicting information, requiring resource-consuming duplicate checks, while conflict-resolution algorithms are lacking.Source: A. Specified Limitations · p. 9
• CBS stores large search and constraint trees in RAM, and the survey warns that computation can outgrow server resources as agent numbers increase.The survey suggests combining reinforcement learning, DPO, and tabu search to record promising paths and avoid deadlocks.Source: A. Specified Limitations · p. 9

B. General Limitations

• Most systems lack standardized specifications, formal syntax and semantics, shared workflow models, compatible execution interfaces, and portable intermediate representations.These gaps make workflows hard to analyze, verify, debug, orchestrate across frameworks, or reuse.Source: B. General Limitations · p. 10

C. Future Direction

• Google's Agent2Agent (A2A) is presented as an open standard for cross-platform agent communication, context sharing, and collaboration.The survey frames this as a promising signal toward industry-wide convergence.Source: C. Future Direction · p. 10
• Future workflows should support context-aware tool use across APIs, search engines, databases, and computational modules.The survey also points to customization for complex, domain-specific tasks, including research and financial analysis.Source: C. Future Direction · p. 10
• Multimodal workflows are expected to handle images, code, documents, and structured data, while multi-agent frameworks advance coordination in decentralized, event-driven settings.Source: C. Future Direction · p. 10
• Unified frameworks and common abstractions are proposed to support composability, interoperability, modular deployment, and long-term agent operation with minimal human input.Source: C. Future Direction · p. 10
• The survey reviews academic frameworks and industrial platforms, comparing systems across functionality and architecture and mechanisms.The comparisons cover 24 agent workflow systems.Source: IX.CONCLUSION · p. 10
• The survey highlights growing needs for standardization, modularity, and orchestration, alongside trends in workflow modeling, execution strategies, and optimization.Source: IX.CONCLUSION · p. 10