From Code to AI Native Apps: The Evolution of Programming Paradigms

Programming Paradigm Evolution

With technological progress, programming paradigms continuously evolve. In the Software 1.0 era, developers programmed computers using languages like Java and Python. In the 2.0 era, neural network parameters are adjusted to program models, and in the large‑model era, prompts become the programming language for AI native applications.

Core Concepts of AI Native Applications

AI native applications treat large language models (LLM) as the programming target, running on GPUs instead of CPUs. Development uses prompts rather than traditional code, creating AI‑native apps that require new thinking about development paradigms.

AI Agent Development Key Issues

Perception: The agent must sense internal and external environments for input and output.

Brain: The large model makes decisions based on learned parameters.

Tool: The agent invokes external tools, such as MCP services.

Memory: Both short‑term and long‑term memory are crucial for context.

AI Native Application Reference Architecture

The architecture centers on AI Agents, which run on compute resources (Kubernetes or function compute) and interact with databases or vector stores. User requests enter via an API gateway, are forwarded to the Agent module, which communicates with models through an AI gateway. Nacos manages dynamic prompts and MCP registration, while the AI gateway handles multi‑model routing, token limiting, and protocol conversion. Asynchronous long‑running tasks use message queues, and all observability data (performance metrics, call chains) are collected via OpenTelemetry and aggregated by LoongSuite.

Spring AI Alibaba

Based on the open‑source Spring AI framework, Spring AI Alibaba adds workflow and Agent support, simplifying AI native app development for Java developers. It also provides higher‑level business scenarios such as Deep Research and Data Agent.

Nacos

Nacos serves as dynamic configuration and service registry, extending to MCP service governance. It enables AI Agents to discover and invoke traditional microservices or private MCP services securely.

Higress

Higress is the core AI gateway, offering LLM caching, vector retrieval, token rate‑limiting, and protocol adaptation for multiple model APIs, as well as MCP proxy capabilities.

Apache RocketMQ

RocketMQ provides a message‑queue mechanism to persist intermediate states of multi‑turn AI Agent interactions, enabling fault‑tolerant resume and reducing costly GPU recomputation.

Observability Solution

Observability addresses three pain points: (1) locating performance bottlenecks, (2) controlling token and cost consumption, and (3) evaluating answer quality. By instrumenting the entire Agent chain with OpenTelemetry probes, metrics such as Token, Error, and Duration (TED) are collected, along with traces that show each step’s latency, token usage, and input/output. Key model‑level metrics include TTFT (time to first token) and TPOT (time per output token). Evaluation combines manual case checks and LLM‑based scoring, feeding results back into continuous improvement loops.

Open‑Source Project Roadmap

Spring AI Alibaba – upcoming A2A protocol support and evaluation console. https://github.com/alibaba/spring-ai-alibaba

Higress – AI plugins and RAG extensions. https://github.com/alibaba/higress

Nacos – dynamic prompt management and A2A support in 3.x. https://github.com/alibaba/nacos

Apache RocketMQ – AI‑specific capabilities to be released soon. https://github.com/apache/rocketmq

LoongSuite – probes for Java, Go, and Python agents, Dify integration, and end‑to‑end observability. https://github.com/alibaba/loongsuite-python-agent