Alibaba's "New Paradigm for AI Agent Development" Collection (PDF file)

AlibabaAIapplication(AI Agent"Developing New Paradigms" is aboutAIApplication development guidance documentation. Detailed introduction.AI Agent(intelligentThe development model, architecture design, core components, and relationship with the large language model (agent)LLMThe document emphasizes the combination method of ) .AI AgentFrom traditional tools tointelligentThe evolution of partners is based onLLMReasoning ability and tool usage to accomplish complex tasksautomaticThe article discusses the construction and execution of enterprise-level MCP services.AIApplication development path selection and cloud-native technologies inAIApplications within applications.AIapplication(AI Agent"Developing New Paradigms" provides technologies such as Function Compute and Cloud-Native API Gateway in...AI AgentBest practices in development help enterprise developers better understand and apply them.AITechnology drives businessintelligentTransformation.

Get Alibaba's "AI AgentDevelopment of New Paradigm CollectionOriginal PDF file, scan the QR code to follow and reply: 20250722

AI Application Overview

• From "tools" to "intelligentThe Evolution of "Partners":AIApplications have evolved from tools that passively execute commands to tools that can interact with the environment, understand intent, and improve themselves.intelligentpartner.
• AI Agent + LLMDual-engine mode:LLMAs the "brain," it is responsible for thinking.AI AgentAs the "hands and feet," they are responsible for execution, and work together.
• Core MCP services for enterprise capabilitiesMCP service standardizes the collaborative relationships among multiple parties and solves the problem of building...AI AgentAddressing the pain points of the skill systemfastDocking.
• BuildAITwo application paths: New development vs. existing system modificationNew development is suitable for disruptive innovation, while existing business transformation is suitable for embedding within existing operations.AIability.

AI Agent Overview

• What isAI Agent:AI Agent yesintelligentThe system, through reasoning, action execution, and continuous self-reflection and error correction using a large language model, can solve complex problems.
• AIapplication core components:AI AgentDepend onLLMThe cycle of reasoning, tool execution, and introspection is used to solve complex problems.
• AI AgentReAct inference mode:AI AgentComplete tasks through analysis, reasoning, action, observation, and introspection.
• AI AgentConstruction patterns andAI AgenttypeThe build patterns include coded and low-code approaches, and the types include auxiliary schemas, standalone products, and auxiliary existing businesses.
• BuildAI Agentcore issuesThe key is to build an enterprise-level MCP management system and choose the right runtime environment.

Function Computation (FC) Overview

• What is Alibaba Cloud Function Compute (FC)?Function Compute is an event-driven, fully managed computing service that allows users to use resources on demand without managing infrastructure.
• Function Compute (FC) is a connector for cloud products.Through its event-driven model, Function Compute can connect to various cloud products to achieve...automaticProcessing procedures.
• Function to calculate FC trigger typeIt supports bidirectional integration, unidirectional integration, and cloud product event triggers to meet different business needs.
• Function Compute (FC) is the glue that holds cloud products together.Provides native triggers for various cloud products to help users.fastEstablish a collaborative workflow.
• Client CasesThis section showcases application examples of Function Compute in different business scenarios.
• Function Compute (FC) Application ScenariosThis section introduces the applications of Function Compute in HTTP applications, content processing, big data, and games.AISpecific applications in scenarios such as...

Function Computation FC asAI Agentruntime

• Function Computation FC SupportAI AgentScheme Topology DiagramThis demonstrates function computation as...AI AgentOverall runtime architecture.
• Function Compute runs as a computing resourceAI Agent– EncodingFunction Compute provides a highly flexible and secure runtime environment that supports multiple languages and frameworks.
• Two types: Running on function computationAI AgentIt includes two types: user-defined and framework-developed.
• Advantages and characteristicsFunction Compute supports trigger mechanisms, on-demand scaling, and dynamic storage mounting, improving resource utilization and flexibility.
• Function Compute builds Chat Al AgentIt solves problems such as execution environment interaction, dependency package uncertainty, and file information path uncertainty.
• Function Compute (FC) + CloudFlow Visualization BuildingAI Agent: Through visualization toolsfastBuildAI AgentIt supports multiple scheduling schemes.
• AIStudio ConsoleProvides visual buildingAI AgentThe platform supports functions such as project management and workflow orchestration.

Function Computation FC asAI Agent Sandbox

• AI AgentSandbox sceneIt introduces different scenarios such as Code Sandbox, Browser Use Sandbox, RL Sandbox and Sim Sandbox.
• Code SandboxUsed to assist in training the coding capabilities of the base model, and supports real-time execution and display of user coding tasks.
• Advantages of Code Sandbox on FCIt supports multiple programming language runtime environments and provides hardware specifications and multi-threaded execution capabilities.
• Browser Use SandboxUsed for online search and auxiliary data collection, supporting Session/Cookie affinity and memory-based expansion.
• Browser Use Sandbox on FC AdvantagesSupports session affinity, optimizes cold start, and solves latency issues.
• RL SandboxUsed forLLMorAI AgentThe algorithm performs targeted reinforcement learning.
• Advantages of RL Sandbox on FCProvide security,High efficiencyHigh efficiency and reproducibility, supporting GPU computing resources.
• Sim SandboxUsed for incarnationintelligentSimulation training, supporting multiple GPU card types.
• Sim Sandbox on FC AdvantagesSupports the Isaac Sim/Isaac Lab environment and has asynchronous task management and simulation training capabilities.

Cloud-native API Gateway &AIGateway Introduction

• Introduction to Cloud-Native API GatewayThe cloud-native API gateway integrates multiple gateway functions and supports unified east-west and north-south traffic management.
• The core role of cloud-native API gateways in application architecture – connecting the ecosystemIt acts as a traffic gateway and API gateway, connecting different applications and services.
• Cloud-native API Gateway – Traffic GatewayIt provides functions such as security protection, traffic protection, and service governance to ensure stable traffic.
• Cloud-native API Gateway – API GatewaySupports API First development mode and provides API protection, monetization and other functions.
• Cloud-native API Gateway – Internal PracticeThe case studies within Alibaba Group demonstrate its stability and performance.

AIGateway ProxyLLMBest Practices

• LLMProblems that customers inevitably encounter in production projectsCustomer atLLMProduction projects face numerous challenges, including high hardware costs, model illusion, single-model risks, compliance audits, performance bottlenecks, and API Key quota limitations.
• AIGateway ProxyLLMScheme S:passAIThe gateway implements functions such as multi-model routing, switching, rate limiting and degradation.
• Cloud-native API gateway proxyLLMThe core benefits of S optionIt solves problems such as cost balancing, model illusion, and switching between multiple models.
• AIGateway ProxyLLMBest PracticesIt provides solutions to problems such as consumer management, high availability of model services, and QPS/Token limitations of closed-source models.

• What is MCP?MCP is a standardized protocol used for connecting...LLMExternal data sources and tools.
• standardizationMCP through standardizationLLMThe way to access external data simplifies the integration of different data sources and tools.
• ModularMCP promotes modular design, allowing for the independent development and maintenance of different components.
• ScalabilityMCP makes adding new data sources or tools easy.SimpleNo major modifications to the existing system are required.
• SecurityMCP provides a structured access model to ensure secure and controlled data interaction.
• Operating mechanism of MCP protocol: Describes user input,LLMReasoning, Client/AgentProcesses such as calling the MCP Server.
• Difference between MCP and Function CallingMCP is a general protocol layer standard, while Function Calling is...Large ModelProprietary capabilities provided by the manufacturer.
• The essence and challenges of MCPThe essence of MCP is a system.Prompt wordsThe combination of collaboration and synergy faces systemic challenges.Prompt wordsChallenges include management, security, and transformation of existing business operations.

AIGateway Proxy MCP Service Best Practices

• AIGateway – MCP Functional Architecture: IntroducedAIThe gateway's MCP functional architecture includes a built-in MCP Server, third-party services, and enterprise services.
• Transform traditional business into an MCP Server with zero code modification – Scenario 1:passAIThe gateway transforms traditional business processes into an MCP Server with zero code, enabling...fastProtocol conversion.
• Solving customer pain pointsSolved the client's developmentAIApplications or existing business operationsAIThe pain points encountered during enhancement.
• fastImplement protocol conversionIt converts traditional services registered in Nacos into MCP services through a gateway, supporting multiple languages and services.
• MCP Registry Definition and CharacteristicsThis section introduces the definition and features of MCP Registry, including RESTful API, health checks, and environment configuration support.
• MSE Nacos as Incremental Value of MCP RegistryMSE Nacos offers additional features such as security management, multiple publishing methods, and version control.
• Transform traditional business into an MCP Server with zero code modification – Scenario 2By using MSE Nacos to dynamically discover and register services, traditional business processes can be transformed with zero code.
• Convert SSE to Streamable HTTP: Convert the MCP-based SSE protocol into a more flexible Streamable HTTP protocol.
• Identity authentication and access control under MCP modeIt provides multiple authentication methods through a plugin mechanism, allowing for flexible management and control of client identity authentication and MCP service access permissions.
• The plug-in mechanism allows for flexible configuration of MCP service proxy strategies.The plugin mechanism supports flexible configuration of MCP service proxy strategies, such as concurrency control and traffic control.

MSE Nacos MCP Registry Best Practices

• Nacosopen sourceCommunity developmentNacos since 2018open sourceSince its inception, it has garnered significant attention and usage, becoming the preferred service registration and configuration center in China, and is widely used by numerous leading enterprises.
• MSE Nacos Applicable ScenariosMSE Nacos is suitable for various scenarios such as service management, traffic scheduling, whitelisting, and application disaster recovery. It supports microservices, high availability, front-end ecosystems, and database applications.AIUnified management and control of MCP Servers in the field.
• AIConfiguration Practices (Nacos: Dynamic Updates)Prompt wordsdata): Achieved via MSE NacosAIIn applicationPrompt wordsData is dynamically updated and supports multiple...AIFrameworks and languages to meet different business needs.
• MSE Nacos 3.0 – MCP RegisterMSE Nacos 3.0 provides an enterprise-grade MCP Registry, supporting functions such as service registration, discovery, health checks, and security management, and integrates with a cloud-native API gateway to achieve MCP protocol conversion.
• Building an Enterprise-Level MCP Registry with MSE Nacos 3.0MSE Nacos 3.0, as a superset of the official MCP Registry, provides enhanced features such as private deployment, advanced retrieval, and security management, solving the problems of multi-MCP service retrieval and token consumption.
• MCP: Security Assurance (Under Planning)The plan will provide MCP market agency, MCP interaction, and MCP Server security guarantees to ensure the security of the MCP ecosystem.
• MCP Effectiveness Verification System – Under PlanningThe plan includes building an MCP Server performance verification system to achieve...AgentIt provides precise trigger verification for MCP tools, offering a view of the results and dynamic adjustment suggestions.

Based on Function Computation (FC)AgentBest practices

• MCP Server on FC reuses high-performance capabilitiesFunction Compute reuses the high-performance capabilities of the cloud-native API gateway, supporting high concurrency and low latency.
• MCP Server on FC Observable System: Achieve observability of function computation through methods such as link tracing and log services.
• Alibaba AgentBest Practices – Hundred Refinements IntegrationThis demonstrates how to integrate Bailian into Function Compute to enable model invocation and task execution.
• Alibaba AgentBest Practices – Inventory CheckintelligentbodyThis section introduces inventory counting.intelligentbodyThe architecture and workflow.
• Alibaba AgentBest Practices – Video Content Recognition and Question AnsweringintelligentbodyDemonstrates video content recognition and question answering.intelligentbodyIts architecture and functionality.
• AWS AgentBest Practices
  • AWS Bedrock AgentBreak down complex tasks into subtasks and use Lambda as an external tool for Function Calling to solve the problem step by step.AWS Bedrock AgentLambda expressions enable step-by-step solutions to complex tasks.
  • AWS achieves this through Bedrock + Lambda.AgentTenant isolation, Lambda asAgentBFF and Function CallingAWS implements this through Bedrock and Lambda.AgentTenant isolation and backend service functions.
  • Azure via OpenAI + Functions provideLLM Best Practices for Application Function CallingAzure via OpenAIand Functions implementationLLM Application Function Calling.
  • Azure enables Multi-functional orchestration through Durable Functions.AgentBuilt on FunctionsAgent/Function Calling Best PracticesAzure implements Multi-Functions through Durable Functions.AgentArrangement andAgentBuild.

  AIApplication of observable systems

  • AIApplication challenges:AIApplications face challenges in areas such as basic resources, model inference, and cost.
  • A typicalAINative application architecture and observability requirements: IntroducedAINative application architecture and observability requirements at each layer.
  • AIFull-stack unified monitoring: Built with PrometheusAIA full-stack monitoring dashboard enables monitoring of model performance, token costs, and GPU resources.
  • Model call end-to-end diagnosticsFull-link tracing of model calls is achieved based on OpenTelemetry Trace.
  • Model generation result evaluationBuild a unified log analysis platform to perform secondary evaluation and analysis on model call logs.
  • Tracing: End-to-end diagnostics for model callsTrace records detailed information about model calls, enabling end-to-end full-link tracing.
  • OrientedLLMDomain-specific Trace semantics of applicationsDefine domain-specific operational semantics, and standardize the storage and visualization of key content.
  • LLMWhat metrics should be considered when applying observability?The document lists the metrics that need to be monitored, such as the number of empty responses, the number of timeouts, and the total timeout factor (TTFT).
  • High-quality data acquisition based on OpenTelemetry: via OpenTelemetry Python AgentAchieve high-quality data collection.
  • Embraceopen source, facingLarge ModelCustomized application:passopen sourceTools and self-developed probesLarge ModelObservability of the application.
  • The Implementation Principle of Non-Intrusive Python Probe TrackingThis paper introduces the implementation principle of non-intrusive Python probe embedding.
  • For streaming scenariosLLM Span segmented data collection and mergingThis feature enables segmented acquisition and merging of streaming data, balancing performance and data analysis needs.
  • Dify Observable Best PracticesIt provides suggestions and observability comparisons for Dify production practices.
  • vLLM/SGLang inference performance observable in practiceTroubleshoot inference performance issues by correlating traces and metrics.
  • based onLLMAchieve model generation resultsautomaticChemical assessment:passLLMPerform model generation resultsautomaticChemical assessment.
  • Model generation result evaluation: Improving the original model with a new modelPromptThe response undergoes a second evaluation to improve the generation effect.

  AIThe impact of new application development paradigms on enterprises

  • Amap's Serverless Business Advertising Platform Practice (API First Architecture)This showcases Gaode's Serverless practices and the advantages of its API-first architecture for advertising services.
  • MCP Server FirstThis section introduces the architecture of MCP Server First and the method of building business processes using a low-code approach.

  Appendix: Core Advantages of Cloud-Native API Gateway Foundation

  • Cloud-native API Gateway – High PerformanceThe cloud-native API gateway outperforms Nginx Ingress in performance and supports hardware acceleration and deep tuning.
  • Cloud-native API Gateway – High Availability (SLA: 99.999%)The cloud-native API gateway provides high availability during the R&D, operation, and change phases.
  • Cloud-native API Gateway – Security CapabilitiesThe cloud-native API gateway provides security features such as consumer authentication, mTLS two-way authentication, login authentication, and traffic protection.
  • Cloud-native API Gateway – Plugin Mechanism (Flexible Expansion)It enables flexible expansion through a plugin mechanism and supports multiple languages and online IDEs.

  Best Practices for Traffic Gateways Appendix

  • Unified Access LayerThe unified access layer solves problems such as network connectivity issues, edge deployment of services, and different protocols.
  • K8s IngressThe cloud-native API gateway supports Kubernetes Ingress to enable services.automaticSynchronization and multi-cluster reuse.
  • End-to-end grayscaleAchieve end-to-end canary deployment through a cloud-native API gateway.
  • Local Multi-ActivityThe same-city multi-active solution achieves multi-availability zone deployment and failover through DNS and MSE governance.

  Get Alibaba's "AI AgentDevelopment of New Paradigm CollectionOriginal PDF file, scan the QR code to follow and reply: 20250722

  WAIC WorldartificialintelligentExhibitor Directory (PDF file)

  How to generate a website using Kouzispace?AIGenerate webpage in one sentence