An agent whose intelligence evolves with every interaction, and whose safety hardens dynamically.
让Agent在交互使用中持续全维度自进化
Full-dimensional self-evolution (OmniEvolve): Skill self-evolution, Context self-evolution, BrainModel self-evolution. Plus Hyper-Harness for safety and Deep Reflexion for task success.
全维度自进化（OmniEvolve）：主动式记忆、Skill自进化、Context自进化、BrainModel自进化。安全动态加固：全新Hyper-Harness与Deep Reflexion双层反思架构。
What Can You Do? 你可以做什么
OmniAgent handles real work across coding, research, ops, and communication. OmniAgent 在编码、调研、运维和多端部署中胜任真实工作。
Coding & Development 编码与开发
Write, debug, and refactor code — the agent learns your project. 编写、调试、重构代码 — Agent 从错误中学习你的项目结构。
Read files, run commands, iterate with reflection retries. The agent builds understanding of your codebase over time and avoids past mistakes. 读取文件、执行命令、通过反思重试迭代。Agent 逐步理解你的代码库，避免重复犯错。
Research & Analysis 调研与分析
Search the web, synthesize findings, cross-reference sources. 搜索网页、综合发现、交叉对比多个信息源。
Fetch and summarize pages, run web searches, and compare results. Proactive Memory remembers what worked and avoids what didn't across sessions. 抓取并摘要页面、执行网页搜索、对比结果。主动式记忆跨会话记住有效方法和已踩的坑。
System Administration 系统运维
Run shell commands safely — four layers of protection guard every operation. 安全执行 Shell 命令 — 四层纵深防护保障每次操作。
LLM review, policy engine, interactive approval, and sandbox protection. Dangerous commands require human consent; SSRF and path traversal are blocked automatically. LLM 智能审查、策略引擎、交互审批、执行沙箱四层防护。危险命令需人工确认；SSRF 和路径遍历在沙箱层自动拦截。
Multi-Channel Deployment 多端部署
Deploy once, reach everywhere: CLI, Web UI, Feishu, Discord, Telegram. 一次部署，全渠道触达：CLI、Web UI、飞书、Discord、Telegram。
WebSocket + HTTP gateway with session management. New channels auto-register via package scanning — just add a Python file and it works. WebSocket + HTTP Gateway，内置会话管理。新渠道通过包扫描自动注册 — 加一个 Python 文件即可接入。
Why Omni Agent? 为什么选择 OmniAgent
Most agent frameworks are stuck at "static instructions + tool calls". OmniAgent goes further. 大多数 Agent 框架停留在"静态指令 + 工具调用"。OmniAgent 更进一步。
| OmniAgent OmniAgent | Typical Frameworks 典型框架 | |
|---|---|---|
| Evolution 进化能力 | OmniEvolve — Skill/Context/BrainModel real-time self-evolution OmniEvolve — Skill/Context/BrainModel 实时自进化 | Static skills / config-dependent 静态技能 / 配置依赖 |
| Execution 执行引擎 | Hyper-Harness — progressive loading, concurrent tools, 4-layer dynamic security Hyper-Harness — 渐进式加载、并发工具、四层动态安全扫描 | Agent loop + static tools 单循环 + 静态工具 |
| Agent Loop Agent 循环 | Deep Reflexion — reflect on failure, extract lessons, retry with insight Deep Reflexion — 反思失败、提取教训、带着洞察重试 | ReAct single loop ReAct 单轮循环 |
| Security 安全防护 | 4-layer: LLM review → policy → approval → sandbox 四层：LLM 审查 → 策略引擎 → 交互审批 → 执行沙箱 | Application-level checks 应用层简单检查 |
| Tool Execution 工具执行 | Parallel with auto-dependency resolution (~3x faster) 自动解析依赖，并行执行 (~3x 提速) | Sequential execution 顺序执行 |
| LLM Backends 模型后端 | 9 providers including local inference (vLLM / SGLang) 9 个提供商，含本地推理 (vLLM / SGLang) | 1-3 providers 1-3 个提供商 |
Three Core Innovations 三大核心创新
The pillars that make OmniAgent different from every other agent framework. 让 OmniAgent 不同于其他 Agent 框架的三大支柱。
OmniEvolve OmniEvolve（全维度自进化）
Intelligence evolves with every interaction, safety hardens dynamically. Agent随交互持续自进化，安全动态加固。
- Proactive Memory — Dual-path alignment of explicit user feedback and implicit LLM induction enables autonomous precipitation and self-evolution of user profiles and memory. 主动式记忆 — 显式用户反馈与隐式LLM归纳的双路对齐，实现用户画像与记忆的自主沉淀与自进化。
- Skill Self-Evolution — Pattern extraction from high-frequency action sequences enables native auto-generation; dual-path feedback enables automatic skill repair. Skill自进化 — 高频Action序列范式提取实现原生自生成；用户交互与LLM诊断双路反馈实现自动诊断与修复。
- Personalization Context — Real-time capture of multi-dimensional preference signals builds adaptive personalized context aligned with individual user preferences. Personalization Context — 多维偏好信号实时捕捉，构建自适应个性化上下文，精准契合用户个性偏好。
- BrainModel Self-Evolution — Online reinforcement learning (GRPO + PRM) feedback loop enables closed-loop model self-evolution during interactive use. BrainModel自进化 — 在线强化学习 (GRPO + PRM) 反馈回路，实现交互中模型闭环自进化。
Hyper-Harness Hyper-Harness（超级脚手架）
Efficient, safe, and intelligent execution scaffold for complex tasks. 高效、安全、智能的执行支架，为复杂任务提供系统支撑。
- Progressive Context Loading — Progressive Disclosure design pattern, loading on demand in graduated stages (L0→L1→L2). 渐进式上下文加载 — 渐进式披露设计模式，按需逐级加载 (L0→L1→L2)。
- Dynamic Tool Execution — Auto-resolves inter-tool dependencies, shifting from serial to async parallel invocation. 动态工具执行 — 自动解析工具间依赖，从串行等待转变为异步并行调用。
- Defense in Depth — Four-layer dynamic security scanning: LLM review → Policy engine → Interactive approval → Execution sandbox. Trust-level classification, unbypassable (industry-first). 纵深防护 — 四层动态安全扫描：LLM审查 → 策略引擎 → 交互审批 → 执行沙箱。信任分级，不可绕过（业内首创）。
- Dynamic Multi-Agent — Sentinel (planning) and Guardian (safety) agents dynamically activated based on task complexity and risk level. 动态多智能体 — Sentinel（规划）与 Guardian（守卫）根据任务复杂度与风险等级动态激活。
Deep Reflexion Deep Reflexion（内外双层反思架构）
Improves task success rate (PASS@1) through dual-layer reflective correction. 通过内外双层反思纠偏，提升任务成功率 (PASS@1)。
- Reflexion Loop — LLM-driven automatic root cause analysis (RCA) and heuristic strategy extraction, dynamically injecting Reflexion into context for closed-loop corrective retry. Reflexion 循环 — LLM驱动的自动根因分析与启发式策略提取，动态注入反思实现闭环纠偏与智能重试。
- Stuck Detection — Three-layer prevention (trajectory repetition, error action repetition, loop pseudo-termination) monitors risks and injects context. 卡死检测 — 三层失败预防（轨迹重复、错误Action重复、Loop伪终止）监测风险并注入上下文。
- Context Compaction — Inner-to-outer failure experience conversion: fault-driven RCA extracts reflection and injects it into the outer loop context space. 上下文压缩 — 内层失败经验转化至外层：故障驱动的自动根因分析提取反思，注入外层上下文空间。
Get Started in 3 Steps 三步快速开始
Install, configure, start chatting. 安装、配置、开始对话。
# 1. Install
$ pip install -e .
# 2. Interactive setup
$ omniagent onboard
# 3. Start
$ omniagent chat # CLI
$ omniagent serve # Web UI → http://127.0.0.1:18790
Switch providers on the fly: omniagent chat --provider deepseek — supports 9 backends including local inference.
随时切换提供商： omniagent chat --provider deepseek — 支持 9 个后端，含本地推理。
Requirements 环境要求
FAQ 常见问题
What is OmniAgent?
OmniAgent 是什么？
OmniAgent is an open-source AI Agent framework built around three core innovations — OmniEvolve (self-evolution), Hyper-Harness (safe execution), and Deep Reflexion (learn from mistakes). It supports 9 LLM providers, 18 built-in tools, multi-channel gateway, and runs on Python 3.11+ with native asyncio.
OmniAgent 是一个开源 AI Agent 框架，围绕三大核心创新构建 — OmniEvolve（全维度自进化）、Hyper-Harness（超级脚手架）、Deep Reflexion（内外双层反思架构）。支持 9 个 LLM 提供商、18 个内置工具、多渠道 Gateway，基于 Python 3.11+ 原生 asyncio 运行。
How is OmniAgent different from other agent frameworks?
OmniAgent 和其他 Agent 框架有什么不同？
Most frameworks use simple ReAct loops with sequential tool execution. OmniAgent adds three dimensions: (1) Self-evolution — the agent writes its own skills and learns from mistakes; (2) Hyper-Harness — parallel tool execution, progressive context loading, and 4-layer security; (3) Deep Reflexion — instead of blind retries, the agent reflects on root causes and retries with insight.
大多数框架使用简单的 ReAct 循环和顺序工具执行。OmniAgent 在三个维度上更进一步：(1) 自进化 — Agent 自动编写技能并从错误中学习；(2) Hyper-Harness — 并行工具执行、渐进式上下文加载、四层安全防护；(3) Deep Reflexion — 不盲目重试，而是分析根因并带着洞察重试。
What LLM providers are supported?
支持哪些 LLM 提供商？
9 backends: DeepSeek, OpenAI, Anthropic, Ollama, Google Gemini, OpenRouter, vLLM, SGLang, and Custom (any OpenAI-compatible endpoint). All support native function calling. You can switch providers on the fly with --provider flag or configure different models per agent.
9 个后端：DeepSeek、OpenAI、Anthropic、Ollama、Google Gemini、OpenRouter、vLLM、SGLang 和 Custom（任何 OpenAI 兼容端点）。全部支持原生 function calling。可以通过 --provider 参数随时切换，或为不同 Agent 配置不同模型。
How does the security system of Hyper-Harness work?
Hyper-Harness的安全系统是怎么工作的？
Four-layer defense: (1) Guardian Agent — LLM-driven contextual safety review; (2) ToolPolicy — rule-based interception with priority engine; (3) ApprovalManager — interactive approval for dangerous commands; (4) Sandbox — SSRF protection, path traversal prevention, dangerous command blocking. All tool executions are audit-logged.
四层纵深防护：(1) Guardian — LLM 驱动的上下文安全审查；(2) ToolPolicy — 基于规则的拦截引擎；(3) ApprovalManager — 危险命令的交互式审批；(4) 沙箱 — SSRF 防护、路径遍历拦截、危险命令阻断。所有工具执行均有审计日志。
What channels are available?
支持哪些渠道？
Terminal CLI, Web UI (served via Gateway with real-time WebSocket), and messaging bots: Feishu, Discord, Telegram. The gateway also exposes a JSON API (POST /message) and health check endpoint. New channels auto-register via package scanning.
终端 CLI、Web UI（通过 Gateway 实时 WebSocket 通信）和消息机器人：飞书、Discord、Telegram。Gateway 同时暴露 JSON API（POST /message）和健康检查端点。新渠道通过包扫描自动注册。
Is OmniAgent open source?
OmniAgent 是开源的吗？
Yes. OmniAgent is fully open source under the GPL-3.0 license. Contributions are welcome.
是的。OmniAgent 基于 GPL-3.0 许可证完全开源，欢迎贡献。