Paper deep dive
Making Sense of AI Agents Hype: Adoption, Architectures, and Takeaways from Practitioners
Ruoyu Su, Matteo Esposito, Roberta Capuano, Rafiullah Omar, June Sallou, Henry Muccini, Davide Taibi
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 4/2/2026, 3:24:52 AM
Summary
This paper presents a large-scale empirical analysis of 138 practitioner conference talks to understand the adoption, architectural strategies, and real-world implementation of AI agentic systems. It identifies key motivations for adoption, recurring architectural patterns (such as reasoning-action-observation loops and supervisor-orchestrated coordination), and the challenges practitioners face when transitioning from experimentation to production.
Entities (4)
Relation Signals (3)
Large Language Models → drives → AI Agents
confidence 100% · These capabilities are largely driven by recent advances in large language models (LLMs)
Practitioners → adopt → Agentic Systems
confidence 95% · To support practitioners in understanding how agentic systems are designed in real-world industrial practice
Model Context Protocol → influences → Agentic Systems
confidence 90% · Third, standardized communication and the maturity of the model context protocol (MCP) influence feasibility.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:To support practitioners in understanding how agentic systems are designed in real-world industrial practice, we present a review of practitioner conference talks on AI agents. We analyzed 138 recorded talks to examine how companies adopt agent-based architectures (Objective 1), identify recurring architectural strategies and patterns (Objective 2), and analyze application domains and technologies used to implement and operate LLM-driven agentic systems (Objective 3).
Tags
Links
- Source: https://arxiv.org/abs/2604.00189v1
- Canonical: https://arxiv.org/abs/2604.00189v1
Trouble viewing inline? Open PDF directly →
Full Text
37,684 characters extracted from source content.
Expand or collapse full text
arXiv:2604.00189v1 [cs.SE] 31 Mar 2026 Article Type: Description (Special Issue on Engineering Agentic Systems) Making Sense of AI Agents Hype: Adoption, Architectures, and Takeaways from Practitioners Ruoyu Su, University of Oulu, Oulu, Finland Matteo Esposito, University of Oulu, Oulu, Finland Roberta Capuano, University of L’Aquila, L’Aquila, Italy Rafiullah Omar, University of L’Aquila, L’Aquila, Italy June Sallou, Wageningen University, Wageningen, The Netherlands Henry Muccini, University of L’Aquila, L’Aquila, Italy Davide Taibi, University of Southern Denmark, Vejle, Denmark & University of Oulu, Oulu, Finland Abstract— To support practitioners in understanding how agentic systems are designed in real-world industrial practice, we present a review of practitioner conference talks on AI agents. We analyzed 138 recorded talks to examine how companies adopt agent-based architectures (Objective 1), identify recurring architectural strategies and patterns (Objective 2), and analyze application domains and technologies used to implement and operate LLM-driven agentic systems (Objective 3). AI agents are rapidly reshaping software engineering practices, supporting activities such as code generation, testing, docu- mentation, workflow orchestration, and system oper- ation. These capabilities are largely driven by recent advances in large language models (LLMs), whose applications continue to be actively explored [1], mak- ing AI agents foundational components of modern AI- driven systems [2], [3]. According to Gartner [4], 40% of enterprise applications will feature task-specific AI agents by 2026, up from less than 5% in 2025. Interna- tional Data Corporation (IDC) [5] also forecasts agent- based AI as a major driver of enterprise IT investment through large-scale agent orchestration. Over the past two years, practitioner demand for concrete guidance and shared experience has in- creased significantly. Newly established conferences on agentic AI report strong participation, while major industry events have expanded their focus on agentic systems. For instance, recent editions of Amazon Web Services (AWS) re:Invent featured 720 sessions on X-X © 2021 IEEE Digital Object Identifier 10.1109/X.0000.0000000 AI agents within a single week, covering architecture, tooling, deployment, and governance [6]. This article adopts a practitioner-centered analysis to identify what works, what fails, and how agent systems are structured in practice. We analyze 234 publicly available practitioner conference talks and technical presentations on AI agents using a structured qualitative approach supported by LLMs with humans in the loop. Rather than proposing new frameworks or formal theories, our goal is to consolidate recurring themes, architectural strategies, and practical lessons reported by practitioners building and deploying agen- tic systems. Specifically, the article is guided by three objec- tives. First, we examine how organizations transition from early experimentation with AI agents to production adoption, focusing on motivations, architectural factors, and practitioner experience influencing whether teams adapt existing systems or build agentic solutions from scratch. Second, we capture reusable architectural knowledge from industrial practice, identifying common strategies and patterns for designing agentic systems. Third, we investigate how agent-based architectures behave in real-world settings, highlighting differences across application domains and the technologies used MonthPublished by the IEEE Computer SocietyPublication Name 1 THEME/FEATURE/DEPARTMENT to implement and operate LLM-driven agentic systems. This article makes three main contributions: • a large-scale empirical analysis of AI agent adoption grounded in 234 practitioner confer- ence talks, providing insight into how organiza- tions transition from experimentation to produc- tion; • a consolidated view of recurring architectural strategies and patterns that characterize agentic systems in industrial practice; • a cross-domain synthesis of agentic system im- plementations, identifying the technologies and frameworks used to build and operate LLM- driven agents. Together, these contributions translate dispersed prac- titioner experience into structured knowledge and prac- tical guidance for engineers designing and deploying agent-based systems. AI AGENTS ARCHITECTURES Agentic systems open new perspectives for design- ing AI agents, from single-agent to multi-agent ap- plications. In the former case, new architecture de- sign decisions are required to: use short-term or long-term memory strategies for storing, organizing, and retrieving context, experiences, or task-relevant knowledge; context compression or summarization for long-running tasks to handle limited context windows; choosing which tools, Application Programming Inter- faces (APIs), plugins, or actuators the agent can use to interact with the world beyond text; optimize for latency and cost by limiting unnecessary tool calls. Those decisions impact the internal structure of an AI agent, including its memory mechanisms, tools, planning and reasoning capabilities, and self-reflection. When designing multi-agent AI systems, decisions on how to divide labor by skill and feedback across agent personas, how to connect agents (e.g., chore- ography or orchestration), and which communication patterns to use become fundamental for engineering quality AI systems. SIDEBAR: The Analysis Process Our study followed a lightweight, practitioner- oriented qualitative analysis process designed to extract architectural insights from a large collec- tion of recorded industry talks and video tran- scripts. The goal was not to build exhaustive theory, but to capture recurring patterns, strate- gies, and challenges communicated by practition- ers when adopting and operating AI agent-based architectures. We applied a recent methodology for applying LLM in qualitative analysis [7], [8]. Data Collection We gathered 234 public practitioner talks, tech- nical briefings, and conference presentations released between September 2024–September 2025. We selected the most relevant practitioner conference using four inclusion criteria: (i) at least five prior editions (≥ 5), (i) talks delivered by practitioners or industry-affiliated researchers, (i) sessions conducted in English, and (iv) full video recordings freely available online [8]. Videos were automatically transcribed and manually checked. An automated pipeline of Bash scripts was used for video transcription, comprising the following steps: i) video download using the open source yt-dlp tool a i) audio extraction and cut using ffm- peg b , and i) transcript extraction using Whisper- XXL c , a Transformer-based deep learning auto- matic speech recognition (ASR) model. Four authors, working in pairs with a final judge resolving disagreements, independently re- viewed the transcripts, applying inclusion and ex- clusion criteria to reduce 234 transcripts to 138. We used a Large Reasoning Model (LRM) as the primary engine for transcript analy- sis, adopting a retrieval-argumented generation (RAG) approach to handle transcripts exceeding the LLM’s context length. The LRM processed each transcript and produced structured outputs that captured architectural statements such as adoption drivers, coordination patterns, workflow structures, and enabling technologies. The LRM’s role mirrors its function in prior architecture com- pliance studies, where it acts as the main rea- soning component for evidence extraction and classification (thematic and axial coding). To ensure accuracy and reduce single-model bias, we employed three independent Validator models (V1, V2, V3). Each Validator reviewed the LRM’s output and checked it against the original transcript segments: • whether the LRM’s summaries faithfully re- flected the transcript, • whether architectural concepts were over- stated or hallucinated, • and whether the extracted insights main- 2 Publication TitleMonth 2021 THEME/FEATURE/DEPARTMENT tained practitioner intent. Validators produced a simple agree/disagree decision with comments. Only insights confirmed by at least two Validators were retained. Finally, according to Su et al. [8] and Ro- bredo et al. [7], two authors reviewed the con- solidated LRM+Validator outputs and assessed the LLM identified themes. Additional details, in- cluding prompts, scripts, extended methodology, and full coding schema, are provided in the online appendix [9]. a https://github.com/yt-dlp/yt-dlp b https://w.ffmpeg.org/ c https://github.com/Purfview/whisper-standalone- win/releases/tag/Faster-Whisper-XXL Research Objectives Organizations exploring AI agent-based architectures face three challenges: understanding adoption, de- signing architectures, and assessing performance after deployment. Our study addresses them through three objectives. Objective 1 examines how companies adopt agent-based architectures, focusing on motivations, the architectural factors, and the practitioner experi- ence that influence whether teams migrate existing systems or build agentic solutions from scratch. Objective 2 captures the reusable architectural knowledge emerging from industrial practice, clarifying how practitioners define agents, assign responsibilities, and organize coordination within and across multi- agent setups. Objective 3 investigates how agent-based archi- tectures are implemented and applied in real-world settings, highlighting differences across application do- mains and the technologies used to build and operate LLM-driven agentic systems. In the following sections, we use Gx to denote a group of talks that provide evidence for a specific claim. Due to space constraints, the full mapping from group IDs to talks is available in the replication package. All in all, the transcripts average 4k words and 23k characters, with moderate lexical diversity (TTR = 0.33, MTLD = 64.73) and relatively stable syntactic complexity (MDD = 2.62), while showing substantial variability and right skewness. Understanding How Companies Adopt Agent-Based Architectures Many teams are experimenting with AI agents, but the path from early pilots to meaningful architectural change remains unclear. Objective 1 examines how organizations progress along this journey. We ana- lyze the motivations for adopting agentic patterns, the architectural factors influencing whether to build from scratch or migrate existing systems, and the practitioner experience, including challenges such as toolchain limitations, engineering complexity, and relia- bility and trust in production. We also analyze adoption over time; Figure 1 summarizes the timeline. Motivation Practitioners consistently point to a mix of technolog- ical and organizational forces driving the adoption of agent-based architectural patterns. Teams are moti- vated by the automation potential enabled by improved reasoning, tool use, and extended autonomous opera- tion, which together reduce manual effort and increase productivity (G1). Cost reductions in frontier model inference make agent-based automation economically competitive with manual processes (G2). In parallel, engineering constraints such as scalability, latency, reliability, and enterprise interoperability make agent architectures an attractive fit for modern industrial systems (G3). Practitioners also emphasize customer- facing benefits, including more personalized, predic- tive, and proactive user experiences (G4). Strategic or- ganizational pressure further supports adoption, as se- nior leadership often encourages teams to restructure products and architectures around intelligent agent capabilities (G5). Architectural factors Talks highlighted seven key architectural considera- tions that influence whether AI agent systems are built from scratch or migrated from existing non-agent ar- chitectures. First, planning, reasoning, and grounding requirements matter. When current systems cannot support agent planning, complex reasoning, search- based action selection, or robust real-world grounding, migration demands heavy refactoring. It often pushes teams toward greenfield development (G6). Second, ease of integration and extensibility can favor migra- tion. Architectures that support incremental evolution allow teams to plug in custom agents via APIs while reusing familiar components, tools, and infrastructure (G7). Third, standardized communication and the ma- turity of the model context protocol (MCP) influence Month 2021Publication Title 3 THEME/FEATURE/DEPARTMENT Pre-20162016–20222023–Early 20252025–2028 Pre-agent foundations Expert systems, rule-based autonomy (G50) Infrastructure build-up Low-latency hardware, early cloud proxies Platform enablers for agent workflows (G51) Framework acceleration GenAI prototypes; programmable agents (WeAgent); RAG mainstream; LandGraph; GPT-4o; early 2025 releases (G52) Year of the Agent and beyond Multi-step invocations; reasoning loops; autonomous workflows Evaluation strengthens; forecast 15% routine tasks by agents (G53) FIGURE 1. Timeline of industrial adoption phases for AI agent-based architectures. feasibility. Mature agent communication protocols and MCP-based tooling make migration more attractive, whereas limited or immature protocol infrastructure can push practitioners to design entirely new agent architectures (G8). Fourth, simplicity in build, test, and release en- gineering lowers migration friction. Clear abstraction layers, separation of concerns, and supervised frame- works help teams evolve existing systems while keep- ing build, test, and release processes stable (G9). Fifth, the balance between workflow control and auton- omy is critical. static, predefined workflows scale well for predictable tasks, but highly autonomous behavior for unpredictable work often requires substantial re- architecting, making a fresh design more appealing (G10). Sixth, the interaction pattern fit plays a role. Mi- gration is easier when current systems already support the required interaction patterns, such as natural lan- guage interfaces and programmatic control; otherwise, bigger architectural changes are needed, which can favor starting from scratch (G11). Finally, unified entity abstraction is a decisive factor. Architectures that treat humans, tools, and language models as a shared type of entity simplify coordination and reasoning. Systems that lack this unified view often require extensive refac- toring, making rebuilding the more pragmatic option (G12). Practitioner Experience Several challenges shape how organizations adopt agents in practice. Many teams expected rapid ac- celeration once agents demonstrated initial value, yet early deployments exposed instability and gaps in toolchains, which highlighted the need for deeper evaluation, monitoring, and framework comparison (G13). End-to-end engineering complexity also posed obstacles. Teams iterated repeatedly on architec- ture designs, encountered integration issues across tools, frameworks, and memory, and faced infrastruc- ture scalability pressures that required coordinated hardware-software evolution and “glue” toolkits for ob- servability and workflow management (G14). Use-case selection emerged as another critical factor. Several agent concepts failed due to weak align- ment with business needs, which pushed practitioners to adopt clearer selection criteria and to combine traditional Machine Learning (ML) and Generative AI (GenAI) when appropriate (G15). Reliability and trust further influenced adoption strategies. Latency, hallu- cinations, and error rates affected user confidence, so teams increasingly relied on staged rollout playbooks that start with simple tasks, gradually expand auton- omy, enforce comprehensive testing, and incorporate continuous production feedback (G16). Architectural Strategies and Patterns Teams building agent-based solutions often ask the same practical questions: How many agents do we need? What responsibilities should each agent take? How should agents coordinate? Objective 2 addresses these questions by identifying reusable architectural strategies. We examine how agents are decomposed, how roles are assigned, and which communication and coordination patterns recur across industrial systems. These strategies and patterns are summarized in Fig- ure 2. Strategies Talks highlighted a diverse set of architectural strate- gies that guide practitioners in structuring and or- 4 Publication TitleMonth 2021 THEME/FEATURE/DEPARTMENT Architectural Strategies and Patterns for AI Agents Architectural Strategies – Task decomposition and role specialization – Supervisor-orchestrated and graph coordination – Incremental growth / agent budgeting – Shared context & unified protocols – Dynamic team composition – Department-aligned roles – Expert ensembles & majority voting – Feedback loops / reinforcement refinement – Tool-centric role design – Minimal core structures – Local–global layered agents Architectural Patterns – Reasoning–action–observation loops – Workflow-based static/dynamic collaboration – Multimodal memory (working, long-term, shared) – Policy-driven safeguards – Message-based asynchronous architectures – Skill registries & protocol routing – Graph-based coordination – Hierarchical, coordinator–executor, teamwork, swarm – Nested dialogues & embedded agents – Centralized coordination control plane FIGURE 2. Compact overview of key architectural strategies and structural patterns identified for Objective 2. ganizing multi-agent AI systems. A common start- ing point is task decomposition and role special- ization, where complex activities are broken into manageable subtasks and assigned to agents with clearly defined responsibilities. Thus, reducing the sys- tem complexity improves execution precision (G17). Many systems rely on supervisor-orchestrated and graph-structured coordination, in which coordinator agents interpret goals, allocate and validate tasks, and manage dependencies through graph-based in- teraction frameworks (G18). Practitioners also adopt incremental growth and agent count budgeting, beginning with a minimal set of agents and adding sub-intelligences only when performance gains are measurable, helping prevent coordination failures and illusion compounding (G19). Collaboration is frequently enabled through shared context and unified com- munication protocols, which ensure agent interop- erability and support scalable teamwork (G20). Some architectures introduce dynamic team composition and role reassignment, where sub-agents can be created or reorganized at runtime to meet workflow needs, quality constraints, or changing environments (G21). Other strategies align agent roles with orga- nizational structures through personal/department- aligned role mapping, matching agents to roles such as sales, HR, or IT to maintain real operational fi- delity (G22). Reliability-driven approaches include ex- pert ensembles and majority decision aggrega- tion, where multiple evaluators operate in parallel and their judgments are combined to improve robust- ness (G23). Several systems employ feedback loops and reinforcement-based self-improvement to allow agents to refine their behavior through iterative correc- tion cycles (G24). In addition, tool-centric role design focuses on building precise toolkits—such as analy- sis, repair, or recovery—rather than increasing agent intelligence, improving controllability, and reducing to- ken consumption (G25). Some systems emphasize a minimal core structure for agents while relying on structured environment support, enabling autonomy without premature rigidity (G26). Finally, multi-level lo- cal–global agent layers divide responsibilities across tiers, using registries and access controls to govern cross-cluster and cross-scope interactions (G27). Patterns Talks identified 10 architectural patterns that shape both single-agent design and multi-agent coordination, organized into four recurring categories: coordination, control, reliability, and evolution. Coordination patterns dominate practitioner discus- sions and address how agents interact and distribute work. A foundational pattern is the iterative rea- soning–action–observation loop, in which agents follow cyclical decision-making processes or struc- tured exploration strategies, such as reconnaissance, enumeration, or knowledge graph construction (G28). Coordination is further supported through workflow- oriented static or dynamic collaboration, enabling agents to execute predefined sequences or to oper- ate autonomously beyond LLM-driven action selection (G29). At scale, practitioners rely on message-based asynchronous communication and microservice deployment, using queues and event-driven interac- tions to support retries, long-running tasks, and dis- tributed execution (G32). More structured coordina- tion emerges through centralized skill and function registries and inter-agent protocol coordination, which enable synchronous or asynchronous capability Month 2021Publication Title 5 THEME/FEATURE/DEPARTMENT discovery and invocation via unified routing and ver- sioned toolchains (G33). Graph-based coordination further models tasks as dependent nodes, enabling precise routing, dependency resolution, and domain- specific dispatching (G34). Across systems, multiple coordination paradigms coexist, including hierarchical supervision, coordinator–executor models, team- work, and swarm or hybrid patterns, each balancing autonomy and oversight in different ways (G35). Control patterns focus on governing agent behavior and the scope of interactions. Nested dialogue struc- tures and embedded agent interactions, such as serialized group chats or agents invoking other agents, allow practitioners to constrain reasoning flows while supporting multi-step coordination (G36). At enterprise scale, a centralized coordination control plane en- forces global policies, manages shared state, supports auditing, and dynamically routes resources, forming the backbone of large multi-agent deployments (G37). Reliability patterns address stability, safety, and compliance. Multimodal memory architectures in- tegrate working, long-term, and collective memory to maintain continuity of reasoning and enable robust cross-agent information sharing (G30). In parallel, policy-driven safeguards and remediation mecha- nisms are introduced to prevent behavioral drift, detect failures, and ensure compliance across agent work- flows (G31). Finally, evolution patterns support long-term adapt- ability. By combining modular coordination mecha- nisms, extensible registries, and layered control in- frastructures, these patterns allow agent systems to evolve incrementally as tasks, environments, and or- ganizational requirements change, without requiring disruptive architectural rewrites. Real-World Implementation and Architectural Implications of Agentic Systems Practitioners want to know what works, for whom, and under which conditions. Objective 3 examines the impacts of agent-based architectures, focusing on key qualities such as scalability, adaptability, and main- tainability. We compare architectural practices across domains and identify the technologies used to build LLM-enabled agentic systems. Application Domain Differences Talks explored several perspectives on how communi- cated architectural practices for AI agents differ across domains and tasks. A first difference concerns in- teroperability setup through discovery and stan- dardized protocols. Domains vary in how agents discover each other and publish capabilities, ranging from machine-readable capability documentation and contextual discovery to stricter use of inter-agent proto- cols that support secure and explainable coordination (G38). A second difference lies in task-execution “interaction contracts”, in which domains shape how agents negotiate task-critical parameters, shared data formats, expected outputs, and the choice be- tween flexible natural-language interaction and precise programming-language interfaces (G39). A third dif- ference involves control-versus-autonomy choices in user experience and workflow design. End-to-end automation often employs tightly controlled, predefined workflows. At the same time, interactive assistants emphasize dynamic and more autonomous flows that remain appropriate for well-bounded tasks rather than highly uncertain, open-ended scenarios (G40). A fourth difference concerns enterprise process encoding as directed acyclic graph (DAG) workflows. Many business settings regulate agent execution using struc- tured, DAG processes in which predefined prompts encode workflow steps, tool integrations, and execution order in accordance with domain-specific constraints (G41). Further differences emerge between digital and physical environments and among specialized do- mains. In digital domains, observe–plan–act imple- mentations often rely on hypertext markup language (HTML) or visual page parsing, manipulation, and API calls, whereas physical robots depend on perception pipelines, semantic scene understanding, and motion control, which leads to substantially different architec- tural realizations across stages of the same meta- architecture (G42). In cybersecurity-focused systems, cybersecurity-specific practices evolve through it- erative trial-and-error. Agents frequently alternate be- tween iterative command-adjustment cycles and struc- tured scenario construction, for example, through re- connaissance and knowledge-graph building, under strong efficiency, cost, and operational constraints (G43). Technologies Talks also described ten categories of technologies used to implement LLM-based agentic systems. A first category is the LLM portfolio and serving runtimes. Industrial systems typically integrate multiple frontier and domain-specific models, such as GPT-4, GPT- 4O, Claude, Gemini, Qwen, DeepSeek, Grok-3 mini, and Forward mini, exposed via embedded endpoints, 6 Publication TitleMonth 2021 THEME/FEATURE/DEPARTMENT inference endpoints, and reorderers with security pro- tections. Some deployments also rely on local or open-source runtimes, including OLAMA, LMStudio, Vectorized Large Language Model Serving System (vLLM), Structured Generation Language (SGLang), and container-based inference stacks (G44). A second category comprises agent frameworks and orches- tration platforms that support multi-agent workflows, planning, tool use, memory, and code generation, us- ing frameworks such as AutoGen, Magentic1, Line Graph and its derivatives, Crew AI, LangGraph, Goose, Land Graph, Agent Development Kit (ADK), Autogens, Pedantic AI, Strands Software Development Kit (SDK), and Amazon Bedrock Agent Core (G45). A third cat- egory is tool use and protocol-based integration. These stacks center on APIs and custom functions, standardized integration standards such as MCP, tool invocation protocols, connection adapters, and related components, including tool servers, API agents, diag- nostic tools, and MCP tool catalogs (G46). A fourth category includes memory, context, vector search, and RAG services, where short- and long-term mem- ory, session history, context stores, knowledge bases, vector databases, and semantic search are combined into RAG-as-a-Service style components (G47). A fifth category focuses on planning methods and work- flow structuring, using React-style loops, chains or trees of reasoning, DAG-based workflows, and pre- execution planning integrated with workflow engines for serialized, dependency-aware execution (G48). Moreover, runtime deployment and distributed execution typically use Kubernetes and microservice architectures, serverless inference, and distributed ex- ecution models, sometimes extended with edge–cloud collaboration, with platforms like Amazon Bedrock Agent Core as deployment backbones (G49). Hard- ware acceleration and specialized compute stacks emphasize dedicated inference hardware, including Graphics Processing Units (GPUs). TAKEAWAYS The following takeaways distill recurring lessons from practitioner experiences into actionable guidance for engineers designing and deploying agentic systems. Treat agent based systems as usable but not self stabilizing technologies. If you are adopting agent based systems today, you should expect uneven reliability and immature evaluation tooling. Stability does not emerge automatically and requires continu- ous attention through testing, monitoring, and cautious rollout practices. Plan for system engineering and integration as the main source of effort. When building agentic systems, you should allocate most of your resources to integrating tools and APIs, designing memory and workflow layers, implementing orchestration, and set- ting up observability and debugging. These engineer- ing tasks dominate development time more than model behavior. Use explicit architectural structure to achieve scalability and maintainability. Agentic systems per- form better when you design them around clear task decomposition, modular roles, shared context, and well defined workflows. Without this structure, many scalability and adaptability benefits fail to materialize. Adapt agentic architectures to the target do- main. You should tailor autonomy, control, and inter- action patterns to your domain constraints. Enterprise workflows, interactive systems, physical robots, and cybersecurity applications each require different archi- tectural choices, and agentic designs do not generalize uniformly across these settings. Build a Control Plane You Can Trust. Standardize your agent integrations with mature protocols (e.g., MCP), then invest in a control plane that gives you rout- ing, monitoring, evaluation, and policy enforcement. Add governance (roles, auditability, and compliance workflows) to deploy and scale agent systems safely, interoperably, and with confidence. LIMITATION While our dataset includes systematically filtered prac- titioner conference talks and technical presentations, it may not capture all architectural practices across the software industry. These sources may emphasize successful or vendor-driven cases and underrepresent smaller or failed efforts. In addition, as our analysis re- lies on LLM-enabled approaches, it may be affected by hallucinations or inaccuracies [10]. Although mitigated through multi-model validation and manual review [8], this risk remains. CONCLUSION AI agents are moving from experimentation to every- day use. Practitioner accounts show how organizations move from pilots to production, driven by automation potential but constrained by reliability, toolchain limi- tations, and engineering complexity. Across industries, recurring coordination strategies and reusable patterns show that successful agentic systems rely on explicit architectural structures, including task decomposition and coordination to achieve scalability and maintain- ability. Differences across domains show that agent- Month 2021Publication Title 7 THEME/FEATURE/DEPARTMENT based architectures do not generalize uniformly, as do- main constraints, interaction patterns, and technology stacks shape implementation choices and operational behavior. As these systems evolve, understanding how to engineer reliable, controllable, and domain-adapted agent architectures remains a key challenge. ONLINE MATERIALS All supplementary materials are available in the online appendix, including transcripts, LLM scripts, inclusion and exclusion criteria, the complete codebook, and ex- tended methodology and linguistic and textual analysis of transcripts [9]. ACKNOWLEDGMENTS This work has been funded by the Research Council of Finland (grants n. 359861 and 349488 - MuFAno) and Business Finland (grant 6GSoft [11]), by FAST, the Finnish Software Engineering Doctoral Research Network, funded by the Ministry of Education and Culture, Finland, and with hardware support by CSC, the Finnish IT Center for Science. Ruoyu Su is a doctoral researcher at the University of Oulu, Finland. He earned his MSc degree in Information Processing Science - Software Engineering from the University of Oulu. His research focuses on the inter- section of Large Language Models, Software Architec- ture, and Empirical Software Engineering. Contact him at ruoyu.su@oulu.fi Matteo Esposito is a postdoctoral researcher at the University of Oulu, Finland, with a European Label PhD from the University of Rome “Tor Vergata”. His research focuses on Generative AI and software architecture, maintenance, quality, and security. Contact him at mat- teo.esposito@oulu.fi Roberta Capuanois a postdoctoral researcher at University of L’Aquila, Italy, where she earned her Ph.D. in Information and Communication Technol- ogy. She is working on quality-driven legacy system modernization using generative AI, with a focus on sustainability–performance trade-offs. Contact her at roberta.capuano@univaq.it. Rafiullah Omar is a postdoctoral researcher at the University of L’Aquila, Italy, where he earned his Ph.D. in Information and Communication Technology. He works on Green AI, large language models (LLMs), and LLM-based agentic systems. He can be contacted at rafiullah.omar@univaq.it. June Sallouis an Assistant Professor in the In- formation Technology Group, at Wageningen Univer- sity, in the Netherlands. Her current research inter- ests include sustainable software engineering, and Green AI. She received her Ph.D. in Computer Science from the University of Rennes, France. Contact her at june.sallou@wur.nl. Henry Muccini is a Professor of Software Engineer- ing in the SWEN research group and leader of the FrAmeLab research laboratory at the University of L’Aquila, Italy. His research primarily focuses on Soft- ware Architecture, Software Engineering and AI, and Green Software Engineering/Green AI. Contact him at henry.muccini@univaq.it. Davide Taibi is LEGO®Chair and Professor of Soft- ware Engineering at the University of Southern Den- mark, 7100 Vejle, Denmark, and a software architecture consultant specializing in cloud-native systems. He can be contacted at davide@taibl.it. REFERENCES 1. G. Franceschelli and M. Musolesi, “On the creativity of large language models,” AI & society, vol. 40, no. 5, p. 3785–3795, 2025. [Online]. Available: https://doi.org/10.1007/s00146-024-02127-3 2. H. Naveed, A. U. Khan, S. Qiu, M. Saqib, S. Anwar, M. Usman, N. Akhtar, N. Barnes, and A. Mian, “A comprehensive overview of large language models,” ACM Transactions on Intelligent Systems and Technology, vol. 16, no. 5, p. 1–72, 2025. [Online]. Available: https://doi.org/10.1145/3744746 3. S. M. Sajjadi Mohammadabadi, B. C. Kara, C.Eyupoglu,C.Uzay,M.S.Tosun,and O. Karaku ̧s, “A survey of large language models: evolution, architectures, adaptation, benchmarking, applications, challenges, and societal implications,” Electronics, vol. 14, no. 18, 2025. [Online]. Available: https://doi.org/10.3390/electronics14183580 4. Gartner, Inc., “Gartner predicts 40% of enterprise apps will feature task-specific ai agents by 2026, up from less than 5% in 2025,” https: //w.gartner.com/en/newsroom/press-releases/ 2025-08-26-gartner-predicts-40-percent-of- enterprise-apps-will-feature-task-specific-ai-agents- 8 Publication TitleMonth 2021 THEME/FEATURE/DEPARTMENT by-2026-up-from-less-than-5-percent-in-2025, Aug. 2025, accessed: 2025-12. 5. International Data Corporation (IDC), “Agentic ai to dominate it budget expansion over the next five years, exceeding 26% of worldwide it spending, and $1.3 trillion in 2029, according to idc,” https://my.idc. com/getdoc.jsp?containerId=prUS53765225,Aug. 2025, accessed: 2025-12;. 6. Amazon Web Services, “Agentic ai sessions at aws:reinvent,”https://registration.awsevents.com/ flow/awsevents/reinvent2025/eventcatalog/page/ eventcatalog?search=&search.areaofinterest= 1744312243120001eo0E, Dec. 2025, accessed: 2025-12;. 7. M. Robredo, M. Esposito, F. Palomba, R. Peñaloza, and V. Lenarduzzi, “What were you thinking? an llm- driven large-scale study of refactoring motivations in open-source projects,” ACM Trans. Softw. Eng. Methodol., Feb. 2026, just Accepted. [Online]. Available: https://doi.org/10.1145/3799717 8. R. Su, N. Ahmad, M. Esposito, A. Janes, D. Taibi, and V. Lenarduzzi, “Emerging trends in software architecture from the practitioner’s perspective: A five-year review,” Journal of Systems andSoftware,vol.236,p.112820,2026. [Online]. Available: https://w.sciencedirect.com/ science/article/pii/S0164121226000543 9. R. Su, M. Esposito, R. Capuano, R. Omar, J. Sallou, H. Muccini, and D. Taibi, “Online materials - making sense of ai agents hype: Adoption, architectures, and takeaways from practitioners,” Jan. 2026. [Online]. Available: https://doi.org/10.5281/zenodo.18154316 10. S. M. S. Mohammadabadi, B. C. Kara, C. Eyupoglu, and O. Karakus, “A survey on hallucination in large language models: Definitions, detection, and mitigation,” 2025. [Online]. Available: https://doi.org/ 10.20944/preprints202510.0540.v1 11. M.A.Akbar,M.Esposito,S.Hyrynsalmi, K. D. Kumar, V. Lcnarduzzi, X. Li, A. Mehraj, T. Mikkonen, S. Moreschini, N. Makitalo, M. Oivo, A.-S. Paavonen, R. Parveen, K. Smolander, R. Su, K. Systa, D. Taibi, N. Yang, Z. Zhang, and M. Zohaib, “6GSoft: Software for Edge- to-Cloud Continuum,” in 2024 50th Euromicro Conference on Software Engineering and Advanced Applications (SEAA).Los Alamitos, CA, USA: IEEE Computer Society, Aug. 2024, p. 499–506. [Online]. Available: https://doi.ieeecomputersociety. org/10.1109/SEAA64295.2024.00082 Month 2021Publication Title 9