自己搭 agentBuild your own agent 12
给程序用的底座与编排框架：agent loop、图工作流、多 agent。Bases and orchestration frameworks for programs: agent loops, graph workflows, multi-agent.
只看这一类 →Show only these →AI 工具AI Tools · Harness
自己搭（进阶）Build your own (advanced)
Agent 运行时 / 编排框架 / SDK 的责任边界、状态持久化能力与权限模型。收录标准：提供 Agent 运行时或编排层、有官方文档可回溯、近 30 天有实质更新。The responsibility boundaries of agent runtimes, orchestration frameworks and SDKs, plus state persistence and permission models. Admission criteria: provides an agent runtime or orchestration layer, has official documentation that can be traced back, and has substantive updates within the last 30 days.
- 这里收录能接任务的产品。它要么在编辑器或终端里替你写代码，要么能被你装进自己的程序，要么能连上你的文件和账号。This site lists products that take on tasks. They either write code for you in your editor or terminal, or can be installed into your own program, or connect to your files and accounts.
- 纯聊天工具不在这里。如果你只想找一个对话问答的网页，去 AI 导航，ChatGPT、豆包、Kimi 这些都在那里。Plain chat products are not here. If you just want a chat webpage, go to the AI Directory — ChatGPT, Doubao and Kimi are listed there.
- 每条档案都是核验快照。字段逐条挂官方出处与核验日期，本站不实测、不排名，未知就写未知。Every entry is a verification snapshot. Each field carries an official source and a checked date; we run no benchmarks, publish no rankings, and write "unknown" when it is unknown.
按任务找Find by task
这一页有 15 个对象。下面的数字是本页真正收录的条数——点进去就是筛选后的结果，不是跨站的总数。This page holds 15 objects. The numbers below are what this page actually carries — clicking filters right here, not a site-wide total.
无人值守跑长任务Run long tasks unattended 3
关掉笔记本它还在跑的：厂商云形态与通用 harness。Keeps running after you shut your laptop: vendor-cloud forms and general harnesses.
只看这一类 →Show only these →先回答一件事：你要不要自己搭Answer one question first: do you need to build one at all
这是本分区存在的唯一理由。如果现成产品已经够用，正确答案是「不要搭」——那就去 成品 Agent，不用往下看。真要自己搭，第一步是认清你要的抽象层。This is the only reason this section exists. If a finished product already does the job, the right answer is "do not build" — go to Finished Agents and stop reading. If you really are building, the first step is being clear which abstraction layer you want.
-
1我要最细的控制权，文件系统我自己管I want the finest control and will manage the filesystem myself原语型（primitives-only）Primitives only只给你 agent loop 的原语，不替你决定文件系统 / 规划 / 记忆怎么做 —— OpenAI Agents SDK 是这一类。You get the raw agent-loop primitives and it does not decide for you how the filesystem, planning or memory work — OpenAI Agents SDK is in this group.
-
2我要一个成熟 CLI 直接变成 API 对象I want a mature CLI turned into an API object包装既有 CLIWrapping an existing CLIClaude Agent SDK 捆绑完整 CLI，默认给全工具无沙箱 —— 你省了造轮子，但权限模型是继承来的。Claude Agent SDK bundles a complete CLI with all tools and no sandbox — you save reinventing it, but the permission model is inherited.
-
3我要开箱即跑的通用长任务助手I want a general long-task assistant that works out of the boxbatteries-includedBatteries includedDeep Agents 官方自述 opinionated：预置更多东西，也替你做了更多决定。Deep Agents describes itself as opinionated: more is preset, and more has been decided for you.
-
4流程要图、要能中断恢复、要多 Agent 协作I need a graph, interruptible resumption, and multi-agent collaboration编排框架：LangGraph / Google ADK / CrewAIOrchestration frameworks: LangGraph / Google ADK / CrewAILangGraph 的落盘粒度最明确（每 superstep 存一次图状态），ADK 的图执行引擎最完整，CrewAI 用「角色分工 + 事件流程」两套抽象。LangGraph has the clearest checkpoint granularity (graph state saved every superstep), ADK has the most complete graph execution engine, CrewAI uses two abstractions: role division and event flow.
-
5我要它常驻、跨平台、能定时跑I want it resident, cross-platform, able to run on a schedule平台型 / 常驻型：OpenHands / Hermes AgentPlatform / resident type: OpenHands / Hermes AgentOpenHands 是带 Web UI 的自托管控制中心；Hermes 更进一步——能挂 Telegram/Discord/Slack/WhatsApp/Signal/Email 六平台，还能 cron 无人值守。OpenHands is a self-hosted control centre with a web UI; Hermes goes further — it attaches to Telegram, Discord, Slack, WhatsApp, Signal and Email, and can run unattended on cron.
-
6我只想要个工具用，不想维护I just want a tool to use, not something to maintain不该来这个分区You should not be in this section这属于「不自建」那一侧，见 成品 Agent。That belongs on the "do not build" side — see Finished Agents.
⚠ 三个最容易踩的坑：「状态」和「上下文」不是一回事——压缩上下文会丢信息，状态持久化才是长期可靠运行的前提（各家在这两层的完备程度差别很大）；「provider 无关」不等于「本地模型可用」（Claude Agent SDK 只支持 Claude，CrewAI 则点名支持 Ollama）；编排框架不给沙箱——权限边界要自己确认。⚠ Three traps people fall into most easily: "state" and "context" are not the same thing — compressing context loses information, while state persistence is what makes long-term reliable operation possible (and vendors differ a lot in how complete these two layers are). "provider-agnostic" does not mean "local models work" — Claude Agent SDK supports only Claude, while CrewAI explicitly names Ollama support. Orchestration frameworks provide no sandbox — you have to establish the permission boundary yourself.
按抽象层分组By abstraction layer
同一层里也能差很远 —— 两个「编程底座」一个是纯原语、一个是全托管包装 CLI，所以先按抽象层分，再看具体对象。Even within a layer the gap is wide — of two "coding bases", one is pure primitives and one is a fully managed CLI wrapper. So group by abstraction first, then look at the object.
编程底座Coding base 5
编程底座Coding base给你 agent loop 或现成 CLI 的编程接口。这层的差别不在功能多少，而在它替你做多少决定 —— 一个只给原语，一个把整套 CLI 连权限模型一起打包给你。A programming interface giving you an agent loop or an existing CLI. The difference here is not how much it does but how many decisions it makes for you — one hands you primitives, the other packages the whole CLI along with its permission model.
Claude Agent SDKClaude Agent SDK
Anthropic
它是什么：Claude Code 的编程接口（Python 版）。 它把一个成熟的 coding CLI 变成 PythonThe Python programming interface to Claude Code — a polished c
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
Codex SDKCodex SDK
OpenAI
「把 Codex CLI 变成 Node 进程里可驱动的对象」。官方 README 首句就是 「Embed the Code"Turns the Codex CLI into an object you can drive inside a Nod
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
OpenAI Agents SDK 是「primitives-only 的多agent harness」。 官方自述 lig"A primitives-only multi-agent harness" — the official descrip
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
Pydantic AIPydantic AI
Pydantic
Pydantic AI 是 Pydantic 出的 Python agent 库。 主打类型安全。官方自述把两个关键词并在一Pydantic AI is the type-safe Python agent loop — official READ
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
smolagentssmolagents
Hugging Face
smolagents 是一个薄 agent 库 —— 它让 agent「用代码思考」。 官方自述是 "barebones lA thin agent library from Hugging Face — it lets agents think
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
编排框架Orchestration 7
编排框架Orchestration把多步骤、多 Agent、要中断恢复的流程显式建成图。给的是控制力，代价是接线与状态管理都归你。Makes multi-step, multi-agent flows with resumable state explicit graphs. It gives you control; the cost is that wiring and state management are yours.
AG2AG2
ag2ai
AG2 是 AutoGen 停更后由社区接棒的续作。 仓库描述直接写着 "AG2 (formerly AutoGen)"。 The community continuation of AutoGen after the original repo
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
CrewAICrewAI
CrewAI Inc
CrewAI 是一个 Python 框架。它用「角色分工」和「业务流程」两套比喻来编排多 agent。 官方自述是 "Fra"Organises multi-agent collaboration through role division and
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
一句话说清它是什么： Google ADK 是「code-first 的 agent 编排框架，核心资产是一个图执行引擎」。In one line: Google ADK is a code-first agent orchestration fr
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
LangGraphLangGraph
LangChain Inc
LangGraph 是一个低层编排框架，专门用来构建有状态的 agent。 官方自述很克制，注意那个 low-level： LangGraph is a low-level orchestration framework for building
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
LlamaIndex FrameworkLlamaIndex Framework
LlamaIndex（run-llama）
一句话说清：LlamaIndex 是一个 Python 库。它把外部数据接进来做检索问答。 ⚠ 但有件事必须先说。 官方已明"A Python library that pulls external data in for retrieval an
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
MastraMastra
Mastra AI
Mastra 是「TypeScript 生态的全栈式 agent 框架」。 装一个包，Agents、图工作流、Memory、Mastra is the batteries-included agent framework for TypeScrip
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
它是什么：「Semantic Kernel 与 AutoGen 合并后的微软官方 agent 底座」。 全名 MicrosoThe Microsoft-official agent foundation that merged Semantic K
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
通用 HarnessGeneral harness 3
通用 HarnessGeneral harness预置了规划、文件系统、记忆等一整套，面向通用长任务。开箱程度最高，可改性最低。Presets planning, filesystem, memory and the rest, aimed at general long-running tasks. Most ready out of the box, least modifiable.
Deep AgentsDeep Agents
LangChain
Deep Agents 是 LangChain 出的通用 agent harness。 它把 filesystem、子代理、Deep Agents is a general-purpose agent harness from LangChain.
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
Hermes AgentHermes Agent
Nous Research
Hermes Agent 是一个常驻的跨平台个人助手，会从经验里自己攒技能。 它不绑在你的笔记本上。 官方 README 说"The only resident personal assistant with a learning loop" —
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
OpenHands Agent CanvasOpenHands Agent Canvas
OpenHands（All-Hands-AI）
它是个自托管的 agent 控制中心。官方 README 现在管它叫 Agent Canvas。 它带的是一整套栈：Web It is a self-hosted control centre for agents — the official R
// line 已是双节点，勿走 esc -->
部分核验Partly verified
详情 →Details →
本站目前是「核验快照」，还没有实测数据This site is currently a verification snapshot — no measured runs yet
下面每一个字段都能回到官方原文核对，但我们还没有在同一批任务上跑过这些工具。Every field below can be traced back to the official source, but we have not yet run these tools on the same batch of tasks. 实测协议已经写好（统一任务、统一验收、记录人工介入与返工次数），任务清单也已就绪，The measurement protocol is written (shared tasks, shared acceptance criteria, recording human interventions and rework counts) and the task list is ready, but it is 尚未执行。not yet run.
所以本站给的是能力边界与证据，不是「哪个更好用」的结论。So what this site gives you is capability boundaries and evidence, not a verdict on which is better to use. 真正的选型请用你自己的输入、预算与验收标准跑一遍。For an actual choice, run them yourself against your own inputs, budget and acceptance criteria.
8 个维度，每格都能回溯8 dimensions, every cell traceable
所有对象在同一坐标系下被描述。All objects are described on the same coordinate system. 点开任一对象，每一格下面都有官方源链接与核验日期——不认同可以自己回去查。Open any object and every cell carries an official source link and a checked date — disagree and verify it yourself.
01
模型与开放条件Model & openness
用谁的模型？能不能换第三方？多模型路由怎么配？Whose model does it use? Can you switch to a third party? How is multi-model routing configured?
02
运行位置Where it runs
本地 / 云端 / 混合？索引在本地还是云端？远程控制不等于云端执行。Local, cloud or hybrid? Is the index local or remote? Remote control is not the same as cloud execution.
03
本地文件Local files
能访问多大范围？未打开的大文件怎么读？符号级能力如何？How wide an area can it reach? How does it read large files you have not opened? How capable is it at the symbol level?
04
关机后的任务Tasks after shutdown
电脑关了还能继续跑吗？这是本地形态最常见的误解来源。Does it keep running once the computer is off? This is the most common misconception about local forms.
05
工具与扩展Tools & extensions
工具链有多宽：MCP · LSP · 插件 · Hooks · 自定义指令。How wide is the toolchain: MCP · LSP · plugins · hooks · custom instructions?
06
上下文与记忆Context & memory
索引算法、上下文窗口、会话恢复与分叉、跨会话记忆边界。Indexing algorithm, context window, session resume and fork, and the boundary of cross-session memory.
07
权限与限制Permissions & limits
沙箱与审批边界、凭据管理、数据是否出境、额度如何计算。Sandbox and approval boundaries, credential handling, whether data leaves your machine, and how quota is charged.
08
适合什么任务Best for
最擅长什么任务？什么情况下别用它？What tasks is it best at? When should you not use it?
全部数据一览All data at a glance
同一张表横向对比。点对象名进详情页看逐格证据。The same table side by side. Click an object name to see per-cell evidence.
| 对象Object | 形态Form | 厂商Vendor | 适合Best for | 核验日Checked | 状态Status |
|---|---|---|---|---|---|
| AGAG2AG2 | harness | ag2ai | 适合。 要用 Python 编排多 agent 会话 / 协议化网络（Hub + channel，含咨询、讨论、工作流适配器） 需要可审计的Orchestrating multi-agent conversations / a protocol-driven network in | 2026-10-08 | partial |
| CAClaude Agent SDKClaude Agent SDK | harness | Anthropic | 一句话：它是一个已经被打磨过的 coding agent。 换模型和自定义 loop 形状，它做不到。 适合： 想要一个已经被打磨过的 coOne line: it is a coding agent that has already been polished. Swappin | 2026-09-30 | partial |
| CXCodex SDKCodex SDK | harness | OpenAI | 适合四类情况： 要把 Codex 塞进自己的 Node 应用 / Electron 应用 / CI 需要非交互执行 需要精确到文件级的权限规A good fit in four cases: You want to put Codex inside your own Node a | 2026-10-01 | partial |
| CACrewAICrewAI | harness | CrewAI Inc | 适合三类场景。 任务能被表述成「一支有角色分工的团队」（Crews） 需要业务逻辑留在普通 Python 里、并显式控制执行路径（FlowsFits: tasks that can be expressed as "a team with defined roles" (Crew | 2026-10-08 | partial |
| DADeep AgentsDeep Agents | harness | LangChain | 适合三类场景。 要一个开箱即用、默认调优好、面向长任务多步流程的通用 Agent。 不想自己拼 filesystem / 记忆 / 子代理。Fits: you want a general agent that works out of the box, is tuned by | 2026-09-30 | partial |
| GAGoogle ADK（Agent Development Kit）Google ADK（Agent Development Kit） | harness | 适合这些情况。 要把多个 agent 组织成有分支、有循环、有人工介入的确定性流程 要用代码优先（code-first）定义 agent 与Fits: organising multiple agents into deterministic flows with branche | 2026-10-01 | partial | |
| HAHermes AgentHermes Agent | harness | Nous Research | 适合这六类场景。 想要一个常驻的跨平台个人助手 （六个平台：Telegram / Discord / Slack / WhatsApp / A fit for these six cases. A resident cross-platform personal assistan | 2026-10-01 | partial |
| LGLangGraphLangGraph | harness | LangChain Inc | 适合四类场景。 agent 流程要跑很久，且必须能从中断处精确恢复 需要人工在流程中间介入并修改状态 要把「工作记忆」与「跨会话长期记忆」分Fits: agent flows that run long and must resume precisely from an inte | 2026-10-01 | partial |
| LILlamaIndex FrameworkLlamaIndex Framework | harness | LlamaIndex（run-llama） | 适合：把外部数据接进来做检索问答。 核心任务是把外部数据源（API / PDF / 文档 / SQL）接进来做检索与问答 需要 300+ 现Fits: bringing external data sources in for retrieval and question ans | 2026-10-01 | partial |
| MAMastraMastra | harness | Mastra AI | 适合： TypeScript / Node 团队要一套「开箱即用」的 agent 框架 —— Agents + 图 Workflows + Fits: TypeScript / Node teams that want an out-of-the-box agent framew | 2026-10-08 | partial |
| MAMicrosoft Agent FrameworkMicrosoft Agent Framework | harness | Microsoft | 它是「图工作流」式的编排框架，适合五类场景。 要把多 agent 组织成有分支、有人工介入、可 checkpoint 恢复的确定性图工作流 It is a graph-workflow orchestration framework — a fit for five cases. | 2026-10-08 | partial |
| OAOpenAI Agents SDKOpenAI Agents SDK | harness | OpenAI | 适合： 想要「薄 harness」，agent loop 原语之外的自己挑。 需要 guardrails 与 human-in-the-loFits: you want a thin harness with built-in guardrails / human-in-the- | 2026-09-30 | partial |
| OHOpenHands Agent CanvasOpenHands Agent Canvas | harness | OpenHands（All-Hands-AI） | 能做的事： 把 agent 放到自己服务器上 7×24 跑 定时或 webhook 触发（Slack / GitHub / Linear /What it is good for: Running agents 7×24 on your own server Scheduled | 2026-10-01 | partial |
| PAPydantic AIPydantic AI | harness | Pydantic | 适合：需要类型安全的 Python 团队，以及要把 agent run 变成 durable 的人。 Python 团队要做「类型安全的结构A fit when: you are a Python team that wants type-safe structured outp | 2026-10-08 | partial |
| SMsmolagentssmolagents | harness | Hugging Face | 适合这四类场景。 认同 code-as-action（官方称比 JSON 工具调用少 30% 步数），并愿意给 LLM 生成代码上沙箱 要 Fits: teams that buy into code-as-action (officially ~30% fewer steps | 2026-10-08 | partial |
为什么这里不给总分排名Why there is no overall ranking
这不是省略，是方法上的明确取舍。This is not an omission but a deliberate methodological choice.
- 缺统一实测，就不给分。跨工具的跑分口径不同——题目、预算、执行环境都不一样，A 的 90 分和 B 的 88 分不可比。No shared measurement, no score. Benchmarks across tools use different rules — different questions, budgets and environments. A 90 and a B 88 are not comparable.
- 「未知」是合法答案。"Unknown" is a legitimate answer.查不到就写「未知」并说明为什么，不用推测填充。缺信息本身也是信息。
- 来源链接、类型与核验日都在详情页里，可以逐条回去复核。The source link, its kind and the checked date are all on the detail page, so you can walk any of them back and re-check.
- lifecycle 单独标注。核验日新不等于数据有效——上游归档后数据会失效，所以单独标
active/maintenance/archived。lifecycle is marked separately. A recent check date does not mean the data is still valid — it goes stale when upstream archives the repo, soactive/maintenance/archivedare tracked separately. - 不同层不混排。成品 agent、自己搭的底座、给 agent 装的工具是三种角色，坐标系不同；硬合成一张 33 行的表只会制造假的可比性。Layers are not mixed. Finished agents, bases you assemble yourself, and tools you install for an agent are three different roles with different coordinate systems; forcing them into one 33-row table would manufacture false comparability.
可信度标记怎么读How to read the confidence marks
详情页与上表都带Both the detail pages and the table above carry confidence，它反映, which reflects 本站数据的完整度how complete this site's data is，不是对产品的评价。, not a judgement of the product.
| verified | 8 个维度均有官方源支撑，且核验日在 90 天内all 8 dimensions have official sources and were checked within 90 days |
|---|---|
| partial | 部分维度标为未知，或官方文档不可访问，或核验日超过 90 天some dimensions are marked unknown, or the official docs were unreachable, or checked over 90 days ago |
| stale | 官方已发布重大变化，本站尚未核验the vendor has shipped a significant change that this site has not re-checked |
本站当前的完整度Current completeness on this site
数据层用 node scripts/audit-gaps.mjs 可随时复核这个数字。
以下是最近一次核验的快照（核验日 2026-09-29）：
| 形态Form | 已补齐Filled | 说明Meaning |
|---|---|---|
| harness | 94/120 | 编排框架的版本演进快、接口变化多，逐条核验成本最高 —— 也是全站最低的一档Orchestration frameworks move fast and change APIs often, so these are the most expensive to verify — and the lowest fill rate on the site |
全站合计 415/446 维度已补齐Site-wide 415/446 dimensions filled（93%）
剩余未补齐项分三类：官方未公开（索引算法、沙箱实现细节本就不对外说明）、需实测才能确定（大仓库表现、CI 无 TTY 行为）、客观渠道不可达。The remaining gaps fall into three kinds: not disclosed by the vendor (indexing algorithms, sandbox internals are simply not documented), requires measurement (behaviour on very large repositories, CI without a TTY), and unreachable through objective channels.
我们选择留白而不是填「已支持」——错误的成本最终由使用者承担。We chose to leave a blank rather than fill in "supported" — the cost of an error is ultimately borne by whoever uses it.
数据来源与复核方式Sources and how to re-check them
全部数据来自开源仓库All data comes from the open-source repository speculcom/ai-agent-guide（CC BY 4.0）。
每个条目都是纯 Markdown + frontmatter，含 8 个维度、证据链Every entry is plain Markdown plus frontmatter, carrying 8 dimensions and an evidence chain接、核验日与可信度标记。仓库内含：
METHODOLOGY.md—— 方法论总纲与四条核心规则axes/—— 每个维度的定义与判定标准（含正反例）SCHEMA.md—— 数据格式规范与构建校验规则tracks/*/tasks/_protocol.md—— 实测协议(尚未执行not yet run，故本站暂无实测结论, so this site has no measured conclusions)
发现错误或有新证据，欢迎提 Issue 或 PR——每条修正都会注明依据与影响范围。Found an error or have new evidence? Open an issue or a PR — each correction notes its basis and blast radius.