🚀 AITurbo Focused on Enterprise AI Innovation | Technical Insights | Practical Implementation Providing in-depth analysis and cutting-edge perspectives for AI applications👆 Click to follow for more AI news
🌟 Introduction: Beyond “Conversation”, the Practical Evolution of AI Agents
Roland Berger, in its report “Five Major Trends in the Generative AI Market in China by 2025”, clearly identifies a core trend: AI Agents are evolving from “conversational interaction” to “task completion”, accompanied by the emergence of more “Phone Use” and “Computer Use” applications. This insight accurately captures the pulse of the market. Industry observations indicate that 2025 is becoming the “Year of AI Agents”, with enterprises decisively shifting their focus from early explorations of AI functionalities to large-scale strategic deployments. This transformation represents not only a gradual technological improvement but also a fundamental change in the purpose of AI within organizations. This article aims to analyze this critical shift from the perspectives of practice, architecture, and business value.
The first wave of generative AI from 2023 to 2024, despite its unsatisfactory performance in directly reshaping core business processes, which even triggered widespread disappointment, is an indispensable preparatory phase. This wave “lays the groundwork for a more transformative second phase (the AI Agent era)”. It has successfully accelerated the overall familiarity of enterprises with AI, helping organizations “establish the necessary capabilities” in prompt engineering, model evaluation, and foundational governance. This initial exploration has built the technical and cognitive “scaffolding” for the upcoming autonomous Agents.
💡 Core Insight: Enterprises are generally disappointed with “chatbots”, which is precisely a strong signal of demand. The industry is gradually realizing that “75% of truly challenging business problems cannot be solved by chatbots”. This disappointment does not stem from a failure of AI technology but rather from a precise clarification of enterprises’ true needs.
Early tools, such as chatbots, are fundamentally limited in that they “retrieve information but do not execute tasks”. It is this significant gap between high familiarity and low process impact that has given rise to clear demands from enterprises. The questions of business managers have shifted from “Can AI understand me?” to “Can AI complete this task for me?“. This shift in questioning is the core proposition that Agentic AI aims to address in 2025.
🚧 The Boundaries of Traditional AI Applications: Why Q&A and Generation Are Insufficient to Reshape Processes
To understand the importance of the current paradigm shift, we must first rigorously examine the boundaries of the first generation of generative AI in enterprise applications. These tools, including enterprise-grade chatbots, Copilot (intelligent co-pilot), and retrieval-augmented generation (RAG) based search systems, are essentially “passive and isolated from enterprise systems”.
The core limitation of these tools lies in the existence of an “execution gap”.
Three Core Limitations:
🔸 Retrieve information but do not execute tasks
Chatbots excel at providing answers but cannot perform actions. They cannot “trigger workflows, update databases, or execute system operations”.
🔸 Isolated from actual workflows
These tools “remain static” and “cannot be integrated into workflows”. This forces human users to become the de facto “integration layer”—employees must manually copy AI-generated content from one window and paste it into another enterprise application (such as CRM, ERP, or spreadsheets) to execute the next steps.
🔸 Fragile and lacking contextual understanding
Traditional chatbots easily “lose context” in “multi-step conversations” and perform poorly in “explaining complex regulations, engineering specifications, or compliance documents”.
The currently popular “Copilot” metaphor itself assumes this limitation. The role of Copilot is to assist the Pilot (the executor), rather than autonomously flying the plane. This model provides local efficiency (for example, writing an email faster) but fails to achieve systemic automation (for example, completely eliminating the business process that requires a human to write emails).
As the analysis points out, most transformations will occur in “automating business processes using LLMs” rather than merely “using chatbots to improve efficiency”. Traditional AI optimized the steps in processes but never truly touched or restructured the workflows themselves. The real process bottlenecks—those multi-step business processes requiring human intervention across applications—still exist.
🔄 The Rise of “Task Closure”: The Emergence of Phone Use and Computer Use
In the face of the clear enterprise pain point of the “execution gap”, the evolution direction of AI Agents must be to achieve “task closure”—that is, moving from “talking” to “executing”. Roland Berger’s analysis report points out the specific forms of this trend’s emergence, namely the rise of “Phone Use” and “Computer Use” applications.
📱 Phone Use: Mobile Execution Across Application Ecosystems
The “Phone Use” Agent signifies that AI is beginning to gain execution capabilities on mobile operating systems. Its core ability is to perform complex “cross-application operations”, no longer limited to the sandbox of a single application.
For example, Zhizhu AI’s AutoGLM 2.0 demonstrates this capability, where users can simply issue a command like “order takeout, book a flight, check housing sources”, and the Agent can autonomously operate “Meituan, JD, Xiaohongshu, and dozens of other high-frequency applications” to complete the entire process. Similarly, vivo’s PhoneGPT (“Blue Heart Little V”) also demonstrates how to autonomously “make calls and book restaurant seats”. In this model, the Agent becomes the user’s full representative on mobile, responsible for handling complex tasks across applications.
💻 Computer Use: Bridging the “Last Stronghold” of Desktop
For enterprises with a large number of complex desktop software and legacy systems, the significance of the “Computer Use” Agent may be even more profound. Such Agents (like Agent S2 or Anthropic’s “Computer Use” concept) are defined as capable of operating graphical user interfaces (GUIs) by “simulating human mouse and keyboard control”.
🔑 Key Technological Breakthrough: The key to this model is that it “relies solely on raw screenshots as input” for operation, no longer requiring “structured accessibility data” (such as the accessibility tree of applications). This “visual-first” approach allows the Agent to locate and operate UI elements (such as buttons, text boxes, menus) through visual understanding.
This trend of “task closure” based on “Phone Use” and “Computer Use” is not just an upgrade of AI capabilities; it represents a direct challenge to the limitations of the traditional API economy. Traditional automation, whether RPA (Robotic Process Automation) or API integration, has fundamental flaws. The “if-this-then-that” logic of RPA is too fragile, as even minor changes in interface elements (like button IDs) can cause scripts to break. API integration is limited by whether enterprises provide clean, comprehensive, and well-maintained APIs—which is a luxury for the vast majority of legacy systems and “long-tail” applications.
The “Computer Use” Agent bypasses this dilemma. By watching the screen like a human, the Agent no longer cares whether an application has an API. It can perform visual localization and operations on “raw screenshots”, making the GUI itself the new, universal API. This capability greatly expands the scope of what can be automated, bringing previously “untouchable” legacy systems that lacked APIs into the realm of intelligent automation for the first time.
🧠 Delving into the Agent Execution Engine: Collaboration of Planner, Executor, and Replanner
How does the AI Agent achieve this autonomous “task closure”? Its technical core is a collaborative architecture known as the “Planner-Executor Framework”. This framework enables the Agent to think and act.
🎯 Planner: The “Brain” of Task Decomposition
The Planner is the “brain” of the Agent. Its responsibility is to receive high-level and often vague goals from users (for example, “help me plan a trip” or “analyze these 15 resumes and generate a report”).
The core action of the Planner is “task decomposition”. It breaks down this complex goal into a clear “multi-step plan” consisting of multiple steps. Each step in this plan should be specific, manageable, and executable sub-tasks. This ability to “thoroughly think through” all steps before execution is a key source of the Agent’s efficiency and cost advantage.
🛠️ Executor: The “Hands” that Call Tools
The Executor is the “hands” of the Agent. It does not concern itself with the user’s original high-level goals but strictly receives the planned steps from the Planner.
The Executor’s responsibility is to execute these specific sub-tasks. It accomplishes this by calling the “tools” within its scope of authority. In the Agent architecture, “tools” is a broad concept that can be an API call, a database query, the execution of a piece of code, or even calling a “Computer Use” Agent to click a certain GUI button.
🔄 Replanner: The “Error Correction Mechanism” for Achieving Closure
The Planner is responsible for planning, and the Executor is responsible for execution. However, for enterprise applications, the most critical component is the Replanner.
The Replanner’s function is to “intervene after each individual task is completed (or fails)”. Its core action is to “adjust the plan based on current progress”. This is the feedback mechanism that achieves “task closure”.
💡 Source of Resilience:
In real enterprise environments, execution is fraught with uncertainty: APIs may time out, files may not exist, web pages may fail to load. When a step executed by the Executor fails, the Replanner will dynamically analyze the error information and formulate a new plan (for example, retry, switch to a backup API, or mark that step as unreachable and continue to the next step). In modern Agent frameworks like LangGraph, this is particularly evident: a conditional node in the workflow “will determine which node to call next”, which may either end the task or loop back to the “Agent” (Executor) node with a new plan after a failure.
The Replanner is a remedy for the fragility of traditional automation tools like RPA. The logic of traditional automation is static; once the preset “if-this-then-that” path is interrupted, the process collapses and calls human engineers to fix the script. This is unacceptable for critical business processes. The Planner-Executor-Replanner architecture introduces a dynamic self-correcting mechanism. The Agent’s workflow is no longer a static script but a dynamic, reasoning-based process. This ability to fail, reflect, and replan in real-time execution is the core distinction between “automation” (which is prone to interruption) and “autonomy” (which is highly resilient). This resilience is the most important technical guarantee that drives enterprises to dare to adopt AI Agents in core processes.
🏗️ Fundamental Restructuring of Enterprise Architecture: Agents Become Execution Engines, Backends Retreat to Governance
The evolution of AI Agent execution capabilities is triggering a “fundamental shift in enterprise software architecture”. The depth of this shift far exceeds the simple “introduction of a new tool”.
According to in-depth analysis by InfoQ, a new architectural paradigm is forming:
🎯 AI Agents Become the New “Execution Engine”: Agents are “transforming from auxiliary tools to operational execution engines”. In traditional architectures, backend systems are responsible for interpreting user intent and orchestrating API calls to execute actions. In the new architecture, the Agent is the new business logic layer. It no longer merely advises backend execution but directly calls services and orchestrates workflows.
⚙️ Traditional Backends “Retreat” to Governance Roles: “Traditional application backends are retreating to governance and permission management roles”. The primary responsibility of the backend is no longer to execute business logic but to govern the Agent that is executing the business logic.
This architecture of “Agents as execution engines” requires a new infrastructure to support and control.
🔌 Standardized Protocol: Model Context Protocol (MCP)
To enable Agents to interact safely and standardly with their “tools” (external systems, APIs, databases), the industry urgently needs a universal standard. The Model Context Protocol (MCP) has emerged. It is likened to the “USB-C port” of AI applications. MCP aims to “standardize the way AI models connect to different data sources and tools”, addressing the “fragmentation and repetitive labor” problem of needing custom integrations for each tool in the past. Rafael Torres, a senior software architect at Expedia Group, positions MCP as a “universal protocol for agent-software system interaction”, its importance comparable to HTTP for the Web.
🛡️ Key Control Point: AI Gateway
As Agents begin to autonomously call APIs, a new type of traffic—”Agentic Traffic”—is quietly exploding. How to manage this outbound traffic initiated by AI? The “AI Gateway” is considered “a missing layer in today’s AI infrastructure”.
The AI Gateway is a “middleware component” or “control point” through which all AI Agent requests to external services must pass. This provides enterprises with a “circuit breaker” to intervene before the Agent goes out of control. For CTOs or architects, the governance function of the AI Gateway is a prerequisite for deploying Agents:
Three Governance Functions of the AI Gateway:
🔐 Security and Auditing: Provides “secure credential handling”, “policy engines”, and a complete “observability and auditing layer” to track every action of the Agent and provide audit logs.
💰 Cost Control: Dynamically selects the lowest-cost model through “routing and cost management” and “rate limiting and quotas” to prevent “out-of-control resource consumption”.
⚖️ Compliance and Data Protection: Implements “output guardrails” and “data privacy enforcement” to automatically filter sensitive content or block private data before the Agent’s response is returned to the user or written to the database.
📊 Gradual Adoption Path for Enterprises
This fundamental shift in architecture is complex and disruptive. Enterprises cannot achieve it overnight. A recommended, pragmatic “three-tier architecture” adoption model (Foundation, Workflow, Autonomous) provides enterprises with a clear evolutionary path:
📌 Foundation Tier: First, establish strict “tool orchestration” and “enterprise security” policies. At this stage, the autonomy of the Agent is strictly limited, focusing on ensuring secure and auditable connections.
🔗 Workflow Tier: After establishing trust, enter the “structured automation” phase. Through prompt chaining, routing, and parallelization, allow the Agent to handle clearly defined, complex multi-step workflows.
🚀 Autonomous Tier: Finally, based on fully established governance, transparency, and trust, evolve to “goal-oriented planning”, allowing the Agent to make autonomous decisions within set boundaries (such as cost and compliance).
💡 Key Insight: The emergence of MCP and AI Gateway is the strongest leading indicator that the Agentic architecture is moving from “laboratory” to “production environment”. In technological development, large-scale innovations (like microservices) are always accompanied by powerful standardization and control layers (like service meshes). We are witnessing the exact same pattern in the AI field: the rise of Agents (innovation) is rapidly followed by MCP (standardization) and AI Gateway (control). This proves that enterprises’ focus has shifted from “Can Agents work?” to “How to make Agents work safely?” In the Agentic era, the true “moat” for enterprises may not be the “smartest” Agent but the most robust, secure, and auditable governance platform.
📈 Measurable Business Value: Breakthroughs in Vertical Industries
This series of complex technological and architectural evolutions must ultimately be proven in terms of business value. The new Agent paradigm is fundamentally changing the cost structure and efficiency boundaries of enterprises, and its value has been measured in multiple vertical industries.
💬 Breakthrough One: Customer Service (From Cost Center to Value Center)
In the customer service field, Agents are evolving from passively answering FAQs to end-to-end solving complex problems.
🔹 H&M Case: The virtual Agent deployed by the company not only achieved 70% of customer inquiries being resolved autonomously, but more importantly, it brought a 25% increase in conversion rates during interactions (such as guiding customers through purchases). This marks a shift in customer service from a pure cost center to a revenue-driving unit.
🔹 Bank of America Case: The virtual assistant “Erica” has completed over 1 billion customer interactions and successfully helped the bank achieve a 17% reduction in call center load. These are quantifiable, significant efficiency gains.
🚚 Breakthrough Two: Supply Chain (From Passive Response to Autonomous Orchestration)
In the highly complex field of supply chain management, the value of Agents lies in compressing the decision cycle of “perception-planning-action” from days to seconds.
Data shows that enterprises applying real-time AI analytics (the precursor to Agents) in supply chains have seen their “decision cycles accelerate by 35%”. The “task closure” Agents take it a step further. They can achieve “autonomous supply chain orchestration”, for example, an Agent can monitor global events 24/7, predict demand changes, automatically replan shipping routes, and dynamically manage inventory. This speed and complexity of decision-making are beyond human teams. Agents can also autonomously execute “automated resource allocation” and “predictive maintenance scheduling”.
🎯 Breakthrough Three: Vertical AI Agents (The Moat of Deep Domain Knowledge)
A notable trend is the rise of specialized “vertical AI Agents”.
The value of these Agents lies in their combination of autonomous execution capabilities with “deep domain expertise”. For example, in healthcare, Agents can assist with “medical coding, treatment summaries, and appointment scheduling”; in the legal field, legal AI Agents represented by “Harvey” can assist with “document review, contract analysis, and regulatory research”. This deep specialization creates a “high barrier to entry”, making their business models more defensible.
💡 Key Conclusion: The ultimate ROI (Return on Investment) of Agents lies not only in cost reduction (which is linear value) but also in extreme compression of decision cycles and new business models (which is exponential value). The example of supply chains is not just about saving labor costs; it is about making optimal decisions at “machine speed” in a rapidly changing market—this is a new capability. McKinsey’s analysis clearly points out the potential for this “new revenue”: for example, “Agents embedded in interconnected products or devices” can “autonomously unlock features or trigger maintenance”, thus achieving “pay-per-use, subscription, or performance-based revenue models”. This is a complete transformation of traditional business models, and such transformation can only be realized through the autonomous execution capabilities of Agents in “task closure”.
🎯 Conclusion: The Maturation of the Intelligent Automation Paradigm
The evolution from “conversational interaction” to “task closure” has become the core issue of enterprise intelligent automation by 2025. The essence of this shift is to reshape AI from a passive Q&A tool into an active task executor.
Technically, this relies on the collaborative framework of Planner, Executor, and Replanner, which equips it with the planning and error-correction capabilities needed to handle complex multi-step tasks.
This capability is driving a fundamental restructuring of enterprise architecture: Agents become the new “execution engines”, while traditional backends retreat to governance roles. To achieve this safely and controllably, standardized protocols like AI Gateway and MCP have become indispensable infrastructure.
Ultimately, this paradigm shift brings measurable business value to enterprises. Whether in achieving automated closure in customer service or autonomous orchestration in supply chains, Agents are compressing decision cycles and reshaping operational models.
💡 Found the content valuable?👍 Like👀 View📤 Share Your support is our motivation for continuous creation Let more people see valuable content