Agent24:让你的个人 AI Agent 24 小时在线、自我进化

Agent24: A 24/7 Online, Self-Evolving Personal AI Agent Framework

Progress-Report #Agent24#AI Agent#Self-Evolving#Electron#Modular Framework#Personal AI#Open Source#AuraAI#Progress-Report
更新于
🇨🇳 中文

📦 项目地址Agent24-Desktop · auraai-packages · Agent24 (Skills)

📐 完整设计文档PLAN.md · decision.md (18 个 ADR) · ROADMAP.md


一个被反复确认的痛点:AI 不会变聪明

你每天用 Claude / ChatGPT / 各种 Agent 工具完成任务:写代码、发文章、查资料、做总结。可是无论用得多熟练,它本身不会变得更懂你

每开一个新会话——所有上下文消散; 每用一个新工具——所有偏好重新教; 每发一篇文章——一样的格式调整重复 N 次; 每排查一次故障——同样的诊断路径走 N 遍。

这不是某个 AI 产品的问题,而是行业默认假设:AI 是即用即抛的工具,不是越用越聪明的伙伴

阿里 AMAP-ML 团队 2026 年 4 月发表的 SkillClaw 论文揭示了同一现象——“Skill 一旦部署就是静态的,每个用户都在重复踩同样的坑”。他们的解法是构建集体进化的 Skill 库:用 Evolver 持续观察轨迹、提炼模式、回写共享 Skill 库。在 WildClawBench 上 Qwen3-Max 经过 6 轮进化,Creative Synthesis 得分相对提升 88.41%。

这给了我们启发,但也让我们意识到:集体进化对企业级管线友好,对”个人 24 小时陪伴的 Agent”还差几层抽象。所以我们启动了 Agent24。


Agent24 是什么

Agent24 是一组让个人 AI Agent 真正 24 小时在线、自我进化的开源工具集。它由三部分组成,对应使用门槛由低到高:

组件给谁用形态
Agent24 (Skills)Claude Code 订阅用户4 个 SKILL.md + 配置模板,安装到 ~/.claude/skills/ 即可用
Agent24-Desktop所有人(含非开发者)跨平台 Electron 桌面应用,承载完整能力模块生态
@auraaihq/* packages模块开发者npm monorepo,提供内核、AI 适配、记忆、进化引擎、能力模块

三层设计的目的:让懂 CLI 的人今天就能用最简版本,让普通用户在 Desktop 落地后即开即用,让开发者通过 npm 包参与生态。


设计理念:四个原则

一、框架与能力解耦

我们最初考虑过”从某个 Electron Agent 应用 fork 后裁掉特定场景代码”。这是错的。正确思路是把 Desktop 做成壳,所有能力变成可插拔的模块

  • 内核(Core)只做演进:Electron 壳、IPC、模块加载、AI 适配层、记忆层
  • 能力(Capability Modules)是可装可卸的 npm 包:@auraaihq/publish-twitter@auraaihq/scrape-rss@auraaihq/module-identity
  • 用户从 Desktop UI 安装/卸载模块 ≈ 后台 pnpm add / pnpm remove

这意味着小红书发布、公众号发布、文件归档、图像处理 ……每一个都是独立 npm 包。框架本身永远精简。

二、框架与 AI 模型解耦

所有业务模块只调用抽象的 AI Layer,不直接依赖任何特定 AI:

模块 → AI Layer → [iDoris (本地、隐私优先) | Claude | OpenAI | LLaVA 本地视觉]

iDoris 是 AuraAI 自研的”个人全景洞察”AI 模型(Prism 启发,见 iDoris 仓库),定位是隐私优先的本地推理。但用户也可以路由到 Claude API 或 OpenAI——这是配置项,不是架构约束。

三、隐私优先的默认设置

借鉴 SkillClaw 的集体进化思路,但默认关闭轨迹外发

  • 所有 ATIF 轨迹、memory、archive 默认仅本地存储(加密 SQLite)
  • 跨设备同步是显式开启项,用 NIP-44 端到端加密通过 Nostr 中转
  • “贡献到社区 SkillBank” 必须用户主动 opt-in,且只发送已蒸馏过的 SKILL.md(不发原始轨迹)

这一条写在 ADR-017 里。

四、自我进化是闭环,不是单点

Agent24 的核心价值在于完成”执行→评估→进化”完整闭环:

  1. 执行:Conversation Layer 接受用户输入 → 任务分解器拆解 → 调度执行(本地/远程模块/其他 agent 协作)
  2. 评估:Phase 3 三阶段评估(Stage 1 correctness gate → Stage 2 全维度 → Stage 3 历史比较)+ 可选外部评估(Codex / agent-speaker / dual)
  3. 进化
    • SkillBank (借鉴 SkillRL):分层 skill 库,每次任务执行前检索相关 skill 注入 context(hot path)
    • Evolver (借鉴 SkillClaw):周期性守护进程扫 ATIF 轨迹,识别 3+ 次重复 pattern,决定 refine 现有 skill 还是 create 新 skill(cold path)
    • Validation gate:Codex MCP 作为校验官,新 skill 通过 review 才合并

这个闭环里,SkillBank 像图书馆(存放检索),Evolver 像编辑部(写新书改旧书)——必须分开,但缺一不可。


架构概览

┌──────────────────────────────────────────────────────────┐
│ 内核 Core(永远存在,不可卸载)                              │
│  Electron Shell · 模块加载器 · IPC · AI Layer 适配器框架    │
│  Memory Layer (L0-L3) · Conversation Layer · 凭据/安全     │
└──────────────────┬───────────────────────────────────────┘
                   │ Module API
   ┌───────────────┼───────────────┬───────────────┐
   ▼               ▼               ▼               ▼
基础模块层       社区模块层       iDoris 能力包装    个人模块层
identity        cos72            input            内容发布
wallet          ├ myshop         process          (publish-*)
comm            ├ mytask         query            信息收集
storage         └ myvote         create           (scrape-*)
shared-memory   communication                     个人助理
ai-bridge       shared-memory                     (files/vision/voice...)

三层模块按”服务对象”分类(不是按”功能”),对应 Mycelium Protocol 的”个人 / 社区 / 城市”三层服务对象:

  • 基础层:身份(AirAccount)、钱包(SuperPaymaster gasless)、通信(agent-speaker / Nostr)、加密存储——任何人都需要的基础设施
  • 社区层:cos72 包含 myshop(积分兑换)、mytask(任务-积分)、myvote(投票)三个社区核心模块——构成完整经济+治理闭环
  • 个人层:内容发布、信息收集、个人助理、学习创作——子模块最丰富、可扩展性最强

详见 ADR-003


技术栈

选型理由
桌面壳Electron 30 + Vite + React 18 + TypeScript跨平台分发成熟、UI 一致性、生态丰富
本地数据better-sqlite3(加密)同步 API、零依赖、性能足够
本地 LLMnode-llama-cpp + GGUF 模型自动检测硬件 + HF mirror fallback + 按需下载
跨平台模型存储metadata-only npm 包 + 运行时下载不污染 npm(详见 ADR-008)
包管理pnpm + workspace + 混合 monorepo单 repo 起步、未来易拆分(ADR-007)
跨 Agent 通信agent-speaker(基于 Nostr)去中心化、NIP-44 加密、无需中心服务器
身份AirAccount + WebAuthn(AAStarTEE 私钥、生物识别登录、无需助记词
支付SuperPaymaster (ERC-4337)gasless 微支付、社区积分代付
未来移动端Tauri 2.0(M5+)包小、性能优、Rust 后端 + Web 前端复用(ADR-018)

为了让 M5 顺利切换到 Tauri 2.0(含 mobile),M0-M4 的设计已经在避免 Electron-only API 和 Node 原生依赖的硬绑定。


路线图(5 个里程碑)

M0 (当前)      仓库与 monorepo 骨架 ✅
               18 个 ADR 决策记录 ✅
               npm scope @auraaihq 注册 ✅

M1 (4-6 周)    内核提取(从 xiaoheishu/desktop 借鉴成熟 Electron 架构)
               第一个真实模块:@auraaihq/publish-blog
               模块接口规格 v0.1

M2 (6-8 周)    后台 daemon + tray icon(永远在线)
               任务分解器 + 调度
               iDoris-SDK 合并入 monorepo(@auraaihq/wechat-bridge)
               第二批 publishers (xiaohongshu / wechat-mp / twitter)

M3 (8-10 周)   Memory L0-L3 完整分层
               @auraaihq/skill-bank + @auraaihq/evolver 落地
               Agent24 (Skills) 迁移为 @auraaihq/skills-* npm 包
               Agent24-Desktop rename → Agent24(去掉 Desktop 后缀)

M4 (10-12 周)  跨用户 skill 共享(opt-in)
               Nostr 分发 skill 更新
               iDoris 主 AI 接入

M5+            模块市场 / 跨设备同步 / 三级 agent 网络(个人 ↔ 组织 ↔ 公共)
               Tauri 2.0 mobile 端

跟同类项目的关系

项目关系Agent24 借鉴的部分
Voyager启发”ever-growing skill library” 的概念
DGM启发 + 已采用results.log archive + 血统追踪
SkillRL启发 + 计划采用分层 SkillBank + 自适应检索
SkillClaw启发 + 计划采用Evolver 的 Refine vs Create 决策 + validated publish gate
MemPalace启发 + 已采用L0-L3 分层记忆 + temporal validity

Agent24 不重复造轮子的部分:底层模型(用 Claude / iDoris / Local)、底层协议(用 Nostr / iLink)、原始 Electron 模板(借鉴 xiaoheishu)。

Agent24 真正在做的事:把以上全部组合到一个面向个人用户、可日常 24/7 运行、模块化可扩展、隐私优先的产品中。


项目状态与参与方式

当前阶段:M0 收尾、M1 启动期。所有架构决策已落定(18 个 ADR 完整记录),代码骨架已就绪。

今天就能用的部分Agent24 Skills — Claude Code 订阅用户安装即用,提供 /evolve/evaluate/setup/org-sync 4 个自进化技能。

# 安装 Agent24 Skills 到 Claude Code
git clone https://github.com/AuraAIHQ/Agent24
cd Agent24 && ./install.sh
# 然后在任意项目里:
claude
> /evolve write a python script to dedupe a CSV

关注后续进展

欢迎参与

  • 提交 issue 讨论你想要的能力模块
  • 在 Agent24 Skills 上提 PR 改进 evolve 流程
  • M1 启动后会开放 module 开发模板,欢迎提交第一批第三方 publisher / scraper

写在最后

我们这一代人见证了 AI 从”问答工具”演化为”工作伙伴”的过程。但绝大多数现有产品仍然是stateless 的接口——它们不记得你昨天教过它什么,不会主动观察你的工作模式,不会把今天的失败转化为明天的 skill。

Agent24 的赌注是:真正属于个人的 AI Agent,必须能自我进化、本地优先、模块化可扩展。这不是一篇论文一个 demo 就能解决的事——它需要框架、需要生态、需要时间。

我们把所有设计决策、备选方案、踩过的坑都公开记录在 decision.md。无论你是对个人 AI 有想法的开发者,还是希望日常工作真正被 AI 增强的用户——都欢迎一起把它建起来。

“让 AI 24 小时为你工作、为你思考、为你进化——而你只需要做你最擅长的事。”

— AuraAI Team, 2026-04-28

📖 相关阅读iDoris 立项思考(Agent24 的底层 AI 模型层)· TurboQuant 在 iDoris 上的可行性分析(KV cache 内存压缩)

🇬🇧 English

📦 Project: Agent24-Desktop · auraai-packages · Agent24 (Skills)

📐 Full Design Docs: PLAN.md · decision.md (18 ADRs) · ROADMAP.md


A Repeatedly Confirmed Pain Point: AI Doesn’t Get Smarter

You use Claude / ChatGPT / various Agent tools every day to write code, publish articles, research, summarize. But no matter how skilled you become — the AI itself doesn’t get to know you better.

Every new conversation: all context lost. Every new tool: all preferences re-taught. Every published article: same formatting tweaks repeated N times. Every troubleshooting session: same diagnostic path walked N times.

This isn’t the fault of any single AI product — it’s the industry’s default assumption: AI is a disposable tool, not a partner that grows wiser with use.

Alibaba’s AMAP-ML team’s April 2026 SkillClaw paper exposed the same phenomenon — “Skills are static once deployed; every user trips on the same rocks.” Their solution: a collectively-evolving skill repository, where an Evolver continuously observes trajectories, distills patterns, and writes back to the shared skill base. On WildClawBench, Qwen3-Max gained 88.41% relative improvement on Creative Synthesis after 6 rounds.

This inspires us, but also reveals: collective evolution works for enterprise pipelines — for “personal AI agents that accompany you 24/7” we need a few more layers of abstraction. So we launched Agent24.


What Agent24 Is

Agent24 is an open-source toolset that makes personal AI agents truly 24/7 online and self-evolving. Three components, ordered by usage barrier (low to high):

ComponentFor WhomForm
Agent24 (Skills)Claude Code subscribers4 SKILL.md files + config templates, drop into ~/.claude/skills/
Agent24-DesktopEveryone (incl. non-developers)Cross-platform Electron desktop app, hosting full capability ecosystem
@auraaihq/* packagesModule developersnpm monorepo: kernel, AI adapters, memory, evolver, capability modules

The three-tier design lets CLI users start today with the minimal version, lets regular users get a turnkey desktop app, lets developers participate via npm packages.


Design Philosophy: Four Principles

1. Decouple Framework from Capabilities

We initially considered “fork an existing Electron Agent app and trim domain-specific code”. That’s wrong. The right approach is to make Desktop a shell, with all capabilities pluggable modules:

  • The Core only evolves: Electron shell, IPC, module loader, AI adapter layer, memory layer
  • Capability Modules are install/uninstall npm packages: @auraaihq/publish-twitter, @auraaihq/scrape-rss, @auraaihq/module-identity, etc.
  • Users install/uninstall modules from Desktop UI ≈ background pnpm add / pnpm remove

This means Xiaohongshu publishing, WeChat publishing, file archiving, image processing… each is an independent npm package. The framework itself stays minimal forever.

2. Decouple Framework from AI Models

All business modules call only an abstract AI Layer, not any specific AI:

Module → AI Layer → [iDoris (local, privacy-first) | Claude | OpenAI | LLaVA local vision]

iDoris is AuraAI’s in-house “personal panoramic insight” AI model (Prism-inspired, see iDoris repo), positioned as privacy-first local inference. But users can also route to Claude API or OpenAI — that’s a config option, not architectural lock-in.

3. Privacy-First Defaults

Borrows SkillClaw’s collective-evolution idea, but defaults to no trajectory leakage:

  • All ATIF trajectories, memory, archives stored locally (encrypted SQLite) by default
  • Cross-device sync is opt-in, NIP-44 end-to-end encrypted via Nostr relay
  • “Contribute to community SkillBank” requires explicit user opt-in, sending only distilled SKILL.md (never raw trajectories)

This is codified in ADR-017.

4. Self-Evolution Is a Loop, Not a Point

Agent24’s core value is the complete “execute → evaluate → evolve” loop:

  1. Execute: Conversation Layer accepts user input → task decomposer → schedule execution (local / remote modules / inter-agent collaboration)
  2. Evaluate: Phase 3 three-stage eval (Stage 1 correctness gate → Stage 2 multi-dim → Stage 3 historical comparison) + optional external eval (Codex / agent-speaker / dual)
  3. Evolve:
    • SkillBank (inspired by SkillRL): layered skill library, retrieves relevant skills before each task to inject context (hot path)
    • Evolver (inspired by SkillClaw): periodic daemon scans ATIF trajectories, identifies 3+ repeating patterns, decides Refine vs Create (cold path)
    • Validation gate: Codex MCP as gatekeeper — new skills merge only after review

In this loop, SkillBank is the library (storage + retrieval), Evolver is the editorial team (write new books, edit old ones) — they must be separate, but neither can be missing.


Architecture Overview

┌──────────────────────────────────────────────────────────┐
│ Core (always present, never uninstallable)               │
│  Electron Shell · Module Loader · IPC · AI Layer Adapter │
│  Memory Layer (L0-L3) · Conversation Layer · Cred/Sec    │
└──────────────────┬───────────────────────────────────────┘
                   │ Module API
   ┌───────────────┼───────────────┬───────────────┐
   ▼               ▼               ▼               ▼
Base Modules    Community        iDoris Wrappers   Personal Modules
identity        cos72            input             content publishing
wallet          ├ myshop         process           (publish-*)
comm            ├ mytask         query             info collection
storage         └ myvote         create            (scrape-*)
shared-memory   communication                      personal assistant
ai-bridge       shared-memory                      (files/vision/voice...)

Three-layer modules categorized by “service object” (not by “function”), aligning with Mycelium Protocol’s “individual / community / city” three-tier service model:

  • Base layer: identity (AirAccount), wallet (SuperPaymaster gasless), comm (agent-speaker / Nostr), encrypted storage — infrastructure everyone needs
  • Community layer: cos72 includes myshop (point exchange), mytask (task-points), myvote (governance) — a complete economy + governance loop
  • Personal layer: content publishing, info collection, personal assistant, learning/creation — most diverse and extensible sub-modules

See ADR-003.


Tech Stack

LayerChoiceRationale
Desktop shellElectron 30 + Vite + React 18 + TypeScriptMature cross-platform distribution, UI consistency, rich ecosystem
Local databetter-sqlite3 (encrypted)Sync API, zero deps, sufficient performance
Local LLMnode-llama-cpp + GGUF modelsAuto hardware detection + HF mirror fallback + on-demand download
Cross-platform model storagemetadata-only npm packages + runtime downloadDon’t pollute npm (see ADR-008)
Package managementpnpm + workspace + hybrid monorepoSingle repo first, easy to split later (ADR-007)
Inter-agent commsagent-speaker (Nostr-based)Decentralized, NIP-44 encrypted, no central server
IdentityAirAccount + WebAuthn (AAStar)TEE private keys, biometric login, no mnemonic
PaymentsSuperPaymaster (ERC-4337)Gasless micropayments, community points sponsorship
Future mobileTauri 2.0 (M5+)Smaller bundle, better perf, Rust backend + reusable web frontend (ADR-018)

To enable smooth migration to Tauri 2.0 (incl. mobile) at M5, the M0-M4 design avoids hard-binding to Electron-only APIs and Node native deps.


Roadmap (5 Milestones)

M0 (now)        Repo + monorepo skeleton ✅
                18 ADR decisions logged ✅
                npm scope @auraaihq registered ✅

M1 (4-6 wks)    Core extraction (borrow from xiaoheishu/desktop's mature Electron arch)
                First real module: @auraaihq/publish-blog
                Module interface spec v0.1

M2 (6-8 wks)    Background daemon + tray icon (always online)
                Task decomposer + scheduler
                iDoris-SDK merged into monorepo (@auraaihq/wechat-bridge)
                Second wave of publishers (xiaohongshu / wechat-mp / twitter)

M3 (8-10 wks)   Memory L0-L3 fully layered
                @auraaihq/skill-bank + @auraaihq/evolver landed
                Agent24 (Skills) migrated to @auraaihq/skills-* npm packages
                Agent24-Desktop renamed → Agent24 (drop "Desktop" suffix)

M4 (10-12 wks)  Cross-user skill sharing (opt-in)
                Nostr distribution of skill updates
                iDoris main AI integration

M5+             Module marketplace / cross-device sync / 3-tier agent network (personal ↔ org ↔ public)
                Tauri 2.0 mobile

Relation to Similar Projects

ProjectRelationWhat Agent24 Borrows
VoyagerInspiration”Ever-growing skill library” concept
DGMInspiration + adoptedresults.log archive + lineage tracking
SkillRLInspiration + plannedLayered SkillBank + adaptive retrieval
SkillClawInspiration + plannedEvolver’s Refine vs Create decision + validated publish gate
MemPalaceInspiration + adoptedL0-L3 layered memory + temporal validity

What Agent24 doesn’t reinvent: underlying models (Claude / iDoris / Local), underlying protocols (Nostr / iLink), original Electron template (borrows from xiaoheishu).

What Agent24 actually builds: combining all the above into a product targeting individual users, daily 24/7 operable, modular and extensible, privacy-first.


Project Status & How to Participate

Current stage: M0 wrap-up, M1 launch. All architecture decisions finalized (18 ADRs full record), code skeleton ready.

Usable today: Agent24 Skills — Claude Code subscribers install and use, providing 4 self-evolving skills: /evolve, /evaluate, /setup, /org-sync.

# Install Agent24 Skills into Claude Code
git clone https://github.com/AuraAIHQ/Agent24
cd Agent24 && ./install.sh
# Then in any project:
claude
> /evolve write a python script to dedupe a CSV

Follow progress:

Welcome contributions:

  • Open issues to discuss capability modules you want
  • Submit PRs to Agent24 Skills to improve the evolve workflow
  • After M1 launch, module dev templates open — first wave of third-party publishers / scrapers welcome

Closing

Our generation has witnessed AI evolve from “Q&A tool” into “work partner”. But most existing products are still stateless interfaces — they don’t remember what you taught them yesterday, don’t actively observe your work patterns, don’t turn today’s failure into tomorrow’s skill.

Agent24’s bet: a personal AI Agent must be self-evolving, local-first, and modularly extensible. This isn’t solvable by a single paper or demo — it needs framework, ecosystem, time.

We’ve publicly logged all design decisions, alternatives, and pitfalls in decision.md. Whether you’re a developer with ideas about personal AI, or a user wanting your daily work truly augmented by AI — welcome to build this together.

“Let AI work for you, think for you, evolve for you — 24 hours a day, while you focus on what you do best.”

— AuraAI Team, 2026-04-28

📖 Related: iDoris Project Launch (Agent24’s underlying AI model layer) · TurboQuant Feasibility for iDoris (KV cache memory compression)

💬 评论与讨论

使用 GitHub 账号登录后发表评论