Agenxia SDK: Innovative Practice of Declarative Agent Runtime and A2A Protocol

Agenxia SDK proposes a declarative agent construction paradigm. It configures workflows via workflow.json, exposes a single A2A endpoint (the start method of JSON-RPC), supports SSE event streams for pushing execution events, uses incremental execution and intelligent caching to optimize performance, achieves A2A protocol interoperability, and lowers the threshold for agent development while maintaining flexibility.

The core concept of Agenxia SDK is the "declarative agent". Unlike traditional imperative agents, developers only need to write workflow.json to describe the workflow diagram. Its core insight is: an agent is essentially a general-purpose workflow executor—conversational interactions, tool calls, and UI component interactions can all be uniformly modeled as workflow execution problems, thus supporting diverse scenarios with a concise API.

The architecture follows the "less is more" principle, exposing only the start method of JSON-RPC as the sole entry point—all interactions are uniformly handled via start. The execution model is efficient: when called for the first time without cache, all nodes are executed on demand; subsequent calls only re-execute the start node and its descendants, while other nodes reuse the cache. For example, during calendar component interactions, only relevant nodes are re-executed, avoiding repeated expensive requests.

The workflow model introduces the concept of ports (Handles), which precisely control data flow forwarding via sourceHandle and targetHandle (e.g., the events field of connector-ics is forwarded to the events port of widget-calendar). The conflict resolution rule is simple: when multiple edges point to the same targetHandle, the last defined edge takes effect, encouraging explicit data flow design.

The module executor is located at modules//execute.js (CommonJS specification), which receives inputs, params, and context parameters and returns outputs. context.llm is optional—the SDK creates a client only when LLM_API_URL and KEY are configured (node configuration takes priority, followed by environment variables). Modules without execute.js default to passing through inputs, which is suitable for pure display UI components.

The /a2a/stream endpoint is provided to return SSE event streams (events like node_start, node_complete), with the event scope limited to the re-executed subgraph for efficient front-end updates. Typical application scenarios include:

1. Conversational agents (sending user messages routed to the agent-core node);
2. Interactive components (e.g., calendar selection triggers execution of relevant nodes);
3. Programmatic workflows (initializing the engine and leveraging cache).

Agenxia implements Google's A2A protocol, supporting agent interoperability and bringing ecological advantages (seamless integration, tools exposed as agents, flexible orchestration). The developer experience is user-friendly: zero-configuration startup (auto-loading relevant files), modular expansion (file conventions), multi-layer configuration (CLI/environment variables/.agent.env), and clear JSON-RPC error messages.

Agenxia SDK represents a paradigm shift in agent development (imperative → declarative, complex APIs → single entry, full execution → incremental execution). It is a lightweight yet fully functional solution, suitable for quickly building AI applications that require interactive components and complex workflows. Its declarative concept and A2A protocol support lay the foundation for the agent ecosystem, which may become the mainstream model in the future.