bigchuidw3/spring-ai-agentx

Spring AI AgentX is a Java framework for building AI agents, built on top of Spring AI and the Reactor library. It provides the components needed to run an agent that can reason through a task, call tools, manage memory across a conversation, and pause to ask a human for input. The design follows the ReAct pattern, where the agent alternates between reasoning about what to do next and acting on that reasoning by calling a tool or producing a response. The core execution engine is reactive rather than graph-based. Instead of defining nodes and edges the way some other agent frameworks do, AgentX drives the agent's multi-turn reasoning loop as a Reactor stream. This makes the agent's output work naturally with asynchronous web frameworks like Spring WebFlux, and allows streaming responses token by token as they are produced. Memory is organized in three layers: short-term memory for the current conversation, a user profile for preferences that persist across sessions, and long-term memory backed by retrieval-augmented search. A separate skills system loads additional instructions or capabilities on demand rather than including everything in the system prompt upfront. Tool dispatch supports both function calling and the MCP protocol. The Human-in-the-Loop mechanism lets the agent pause at a defined point and wait for a user's answer before continuing. The framework supports several AI model providers including Qwen, DeepSeek, GLM, and MiniMax, with built-in handling for the differences in how each model returns its chain-of-thought output. A custom model adapter for DeepSeek V4 addresses an incompatibility with Spring AI's standard interface for that model's reasoning output format. The framework requires JDK 21 or later, Spring Boot 3.5.x, and Spring AI 1.1.0.