Zing Forum

Reading

Agent4j: A Pure Java-Implemented AI Programming Agent Framework

Agent4j is a Java 17-based AI coding agent framework benchmarked against Claude Code and Devin. It enables large language models to independently understand, write, and debug code through reasoning loops, tool calls, and streaming output. It supports three deployment forms (CLI, Web, Electron desktop) and leverages prefix caching technology to reduce API costs.

AI 编程代理JavaLLMClaude CodeDevin代码生成自动化开发前缀缓存MCP
Published 2026-06-15 20:16Recent activity 2026-06-15 20:21Estimated read 9 min
Agent4j: A Pure Java-Implemented AI Programming Agent Framework
1

Section 01

Agent4j: Guide to the Pure Java-Implemented AI Programming Agent Framework

Agent4j is a pure Java17-implemented AI programming agent framework maintained by ezdemo, benchmarked against Claude Code and Devin. It was released on GitHub on June 15, 2026 (project link: https://github.com/ezdemo/agent4j). Through reasoning loops, tool calls, and streaming output, it allows large language models to independently complete code understanding, writing, and debugging. It supports CLI, Web, and Electron desktop deployment, and uses prefix caching technology to reduce API costs, filling the gap of AI programming agents in the Java ecosystem.

2

Section 02

Background: The Gap of AI Programming Agents in the Java Ecosystem

Current mainstream AI programming agents (such as Claude Code, Codex, and Devin competitors) are mostly built based on TypeScript, Go, or Rust. Enterprises and developers deeply engaged in the Java ecosystem face high integration and maintenance costs. The emergence of Agent4j solves this pain point, providing a native AI programming assistant for Java tech stack teams.

3

Section 03

Core Architecture and Tool System

Core Architecture: Autonomous Execution Driven by Reasoning Loop

Agent4j's core is the reasoning loop: User input → LLM thinking → Tool call → Observe results → LLM rethinking → ... → Task completion. This loop includes optimizations: context folding (intelligent summarization of old messages), message self-healing (fixing truncated JSON/tool_calls), circuit breaker (preventing infinite loops), and streaming output (SSE real-time push).

Tool System: Expanding Capability Boundaries

Built-in tools cover the entire development process: file operations (read/write/edit), code search (glob/grep/codesearch), command execution (bash), network operations (webfetch/call_api), memory system (remember/recall_memory), plan management (submit_plan/revise_plan), user interaction (ask_choice), and background jobs (run_background). Extension methods: declarative tool development (inherit from AgentTool base class), MCP protocol support, OpenAPI integration, and a built-in skill market for users to install community toolkits.

4

Section 04

Multi-Form Deployment and Cost Optimization

Multi-Form Deployment

  • CLI: Used in the terminal, supports slash commands (e.g., /new to create a session, /plan for plan mode, /execute to run, /compact to fold context, etc.).
  • Web: Based on Vue3 + Vite + Ant Design Vue, providing visual dialogue, tool call display, Git panel, multiple themes, and code highlighting.
  • Desktop: Built with Electron, combining web interface and native experience.

Cost Optimization: Prefix Caching

Prefix caching uses LLM server-side KV Cache to hit repeated system prompts, tool definitions, etc., reducing input token costs. It has significant optimizations for DeepSeek (hit rate ≥97%, cost reduced to 3%) and Xiaomi Mimo (hit rate ≥98%, cost reduced to 2%) models.

5

Section 05

Security and Parallel Processing Mechanisms

SubAgent

Supports creating sub-agents to process independent tasks in parallel. Features: isolated execution (independent context/reasoning loop), tool inheritance (excluding recursive spawn), independent channels (event stream push), and usage statistics (token consumption statistics by model).

Human-in-the-Loop (HITL) Approval

Provides an approval mechanism for write operations: waits for human approval/rejection before execution, supports /agree (approve) and /deny (reject) commands; configurable default approval mode, with editMode options: auto (needs confirmation) and yolo (execute directly).

6

Section 06

Project Structure and Tech Stack

Project Structure

Uses Maven multi-module: agent4j/ ├── agent4j-tool/ (tool abstraction layer) ├── agent4j-bin/ (core engine: reasoning loop, session management, MCP) ├── agent4j-web/ (Web backend: REST interface, SSE push) ├── agent4j-front/ (Vue3 frontend + Electron desktop) ├── intro/ (official website) └── docs/ (documentation)

Tech Stack

  • Backend: Solon4.0.0-M3, Snack4, OkHttp
  • Frontend: Vue3.4, Vite5, Ant Design Vue4
  • Desktop: Electron + Node.js
  • Persistence: JSON Lines
  • Documentation: Knife4j
7

Section 07

Conclusion and Selection Recommendations

Benchmark Analysis

Product Implementation Language Features
Claude Code TypeScript Feature-rich, mature ecosystem, closed-source
Codex TypeScript Official OpenAI product, deeply integrated with GPT
OpenCode Go Lightweight and fast, open-source
Reasonix Rust High performance, memory-safe
Agent4j Java17 Native to Java ecosystem, easy to integrate, open-source

Selection Recommendations

Suitable scenarios: Java tech stack teams, enterprise intranet deployment, need for deep customization, cost-sensitive scenarios.

Quick Start

  • Windows (PowerShell): irm https://raw.giteeusercontent.com/ezdemo/agent4j/raw/main/.release/setup.ps1 | iex
  • macOS/Linux: curl -fsSL https://raw.giteeusercontent.com/ezdemo/agent4j/raw/main/.release/setup.sh | bash Start Web service: agent4j web 0 (First launch requires configuring the API Key in ~/.agent4j/config.json).

Conclusion

Agent4j is an important step for AI programming agents in the Java ecosystem, providing enterprise-level applications with capabilities comparable to mainstream products. For Java developers, there's no need to step out of the familiar tech stack to experience the efficiency improvement of AI autonomous programming—it's worth trying.