CrewAI
面向实际试用和选型：CrewAI is an agent framework for building multi-agent workflows with roles, tasks, tools, memory, and orchestration patterns.
1300
AI Agents
32 agents available
Allagent-framework (1)agent-network (1)agent-skills (4)analysis (3)autonomous-agent (1)browser-agent (2)business (1)coding (3)coding-agent (9)marketing (2)open-source-agent (1)personal-agent (2)productivity (1)security (1)
agent-framework
Cline
面向实际试用和选型：Cline is an open-source autonomous coding agent for VS Code-style workflows, capable of editing files, running terminal commands, using browser tools, and working with MCP.
1000
Aider
面向实际试用和选型：Aider is an open-source AI pair-programming and coding-agent tool that works in the terminal, edits local repositories, and integrates with Git workflows.
1900
Skyvern
面向实际试用和选型：Skyvern is an AI browser automation platform for automating web workflows with computer vision, browser actions, and agentic task execution.
700
browser-use
面向实际试用和选型：browser-use is an open-source browser automation agent framework that lets LLMs operate websites through browser actions for research, QA, and workflow automation.
600
OpenHands
面向实际试用和选型：OpenHands is an open-source software-development agent platform for building, running, and customizing coding agents that can edit code, run commands, and browse.
2100
Devin
面向实际试用和选型：Devin by Cognition is an autonomous AI software engineer designed to take software tasks, work through implementation steps, and return tested changes.
1400
Cursor Background Agents
面向实际试用和选型：Cursor Background Agents extend Cursor's AI editor workflow into asynchronous repository tasks while keeping developers inside the IDE-centered coding experience.
1200
GitHub Copilot Coding Agent
面向实际试用和选型：GitHub Copilot coding agent lets developers assign tasks from GitHub issues and pull requests so the agent can work inside GitHub's review and branch workflow.
1000
Google Jules
面向实际试用和选型：Google Jules is Google's autonomous coding agent for asynchronous software tasks, running work in cloud virtual machines and returning verified changes.
1200
OpenAI Codex
面向实际试用和选型：OpenAI Codex is OpenAI's software-engineering agent family spanning Codex CLI, IDE, desktop, and cloud task workflows for coding, bug fixing, and code review.
800
Claude Code
面向实际试用和选型：Claude Code is Anthropic's terminal-native coding agent for reading repositories, editing files, running commands, using MCP tools, and carrying software tasks through a plan-edit-test loop.
800
ClawSecure OpenClaw Security
面向实际试用和选型：ClawSecure OpenClaw Security tracks security issues around OpenClaw-style agents, especially skill supply chain risk, exposed instances, prompt injection, and unsafe permissions.
700
Moltbook
面向实际试用和选型：Moltbook is an AI-agent social network associated with the OpenClaw ecosystem, notable as an experiment in agent-to-agent communication, identity, and autonomous posting.
1000
Clawdbot / Moltbot
面向实际试用和选型：Clawdbot and Moltbot are historical names tied to the OpenClaw personal AI assistant ecosystem, useful for users searching older tutorials, repos, and community posts.
700
Awesome OpenClaw Skills
面向实际试用和选型：Awesome OpenClaw Skills is a curated GitHub list for discovering OpenClaw-compatible skills across coding, DevOps, browser automation, AI, research, notes, and productivity.
600
Oh My OpenClaw
面向实际试用和选型：Oh My OpenClaw is a community resource hub for finding and installing OpenClaw, Moltbot, and Clawdbot skills and workflow extensions.
1200
OpenClaw Skills Directory
面向实际试用和选型：OpenClaw Skills Directory is a discovery surface for browsing the fast-growing OpenClaw skill ecosystem by category, task type, and install command.
800
ClawHub
面向实际试用和选型：ClawHub is the public registry for OpenClaw skills and plugins, letting users publish, version, discover, and install reusable agent capabilities.
1200
Claw Code
面向实际试用和选型：Claw Code is an open-source AI coding agent framework described as a clean-room Python and Rust rewrite of Claude Code-style agent harness architecture.
1200
Hermes Agent
面向实际试用和选型：Hermes Agent is Nous Research's open-source autonomous agent focused on persistent memory, local infrastructure, and model-flexible long-running assistance.
1300
OpenClaw
面向实际试用和选型：OpenClaw is a viral open-source personal AI assistant that runs as a self-hosted, always-on agent across chat apps, local tools, memory, skills, and automation workflows.
1200
Agent Safety Reviewer Agent
面向 2026 年 AI 工作流的专业内容：Reviews agent workflows for prompt injection, excessive autonomy, secret leakage, and unsafe tool permissions.
500
Community Signal Analyst Agent
面向 2026 年 AI 工作流的专业内容：Synthesizes X, Reddit, GitHub, and Product Hunt signals without over-trusting viral claims.
800
AI Agent 选型
评价 AI Agent，不能只看回答是否流畅，要看执行是否可控。
Agent 市场正在从聊天助手走向真正的运行时：能读取文件、调用工具、写入产物、保存状态并恢复任务。Agent 目录应该帮助用户比较自治程度、权限边界、可观察性、记忆机制和部署适配，而不是只写“提升效率”。
真实变化
看清能力边界，再决定是否采用。
沙箱工作区
Agent 需要一个受控的执行环境
OpenAI 在 2026 年 4 月 15 日更新 Agents SDK，重点支持 Agent 在受控沙箱中查看文件、运行命令、编辑代码并执行长任务。因此，工作区设计、文件权限和输出目录已经成为 Agent 评估标准。
从 Demo 到 Runtime
生产级 Agent 需要记忆、工具、轨迹和恢复能力
OpenAI 提到 configurable memory、MCP、skills、AGENTS.md、shell、apply_patch、snapshot 和 rehydration；Google ADK 也把 tools、sessions、memory、observability、evaluation 和 safety 放在核心位置。
可靠性诊断
很多 Agent 失败，本质是上下文失败
LangChain 文档指出，Agent 经常不是因为模型不够强，而是因为没有给模型正确上下文。用户选择 Agent 时，应重点看它如何选择工具、保存记忆、过滤任务上下文。
热门入口
按搭建、自动化和平台选择 AI Agent
这里可以按目标选择 AI Agent：了解 AI agent 基础、学习 how to create an ai agent、连接 n8n AI agent 自动化、比较 OpenAI Agent 和 Google AI Agent，或评估 Vertex AI Agent Builder。
ai agent
定义与目录发现适合先理解 AI agent、AI agents、AI agent examples、AI agent tools 和典型业务场景。
how to create an ai agent
搭建教程适合需要搭建路径的用户：目标定义、工具、记忆、审批、沙箱、部署检查和评估标准。
n8n ai agent
自动化工作流适合自动化用户，把 AI agents 和触发器、应用集成、审批、工作流和重复运营任务连接起来。
按目标探索 AI Agents
按使用目标快速筛选。
页面需要同时满足初学者、构建者、平台比较者和自动化用户。
构建型意图
这类词需要实际架构说明：目标、工具、记忆、runtime、审批、可观察性、评估和上线标准。
how to build an ai agenthow to create an ai agentopen ai agent buildervertex ai agent builder
平台型意图
用户在比较生态。文案可中立覆盖 OpenAI Agent、Google AI Agent、Manus AI Agent、Replit AI Agent 等平台型搜索。
openai agentgoogle ai agentmanus ai agentreplit ai agent
自动化与学习意图
这些用户需要示例和学习路径。页面可以把 AI agents 与自动化流程、课程和实际搭建联系起来。
n8n ai agentai agent n8nai agents courseai agents intensive course with google
Agent 类型对比
从能力、成本和风险做选型。
轻量助手和生产级 Agent 不是同一种产品。比较 Agent 时，应先看它被允许做什么。
Assistant Agent
- 在聊天流里回答和起草内容
- 集成风险低
- 适合写作、问答、头脑风暴和摘要
Workflow Agent
- 调用工具并在系统之间移动数据
- 需要鉴权、重试和审批节点
- 适合研究、客服、报告和运营
Sandbox Agent
- 在文件、代码、终端或隔离工作区内工作
- 需要严格权限和审计日志
- 适合编程、数据室、文档处理和长任务
使用 Agent 前要检查什么
把工具放进真实工作流。
工具权限：它可以读取、写入、删除、发布或发送什么？
记忆模型：哪些内容会跨会话保存，用户能否查看或清除？
人工审批：哪些动作必须暂停并等待确认？
可观察性：能否看到工具调用、中间产物、失败原因和最终证据？
相关 AI Agent 主题
从这些主题可以继续进入 Agent builder、自动化平台、开源 Agent、免费 AI Agent 和课程路线。
ai agentai agentsopenai agentgoogle ai agentn8n ai agentvertex ai agent builderhow to build an ai agentai agents vs agentic aiai agent freegemini cli open source ai agent
Agent FAQ
常见问题与选型答案。
什么样的 Agent 才算生产级？
它需要有清晰任务边界、受控工具、记忆规则、风险动作审批、日志或轨迹、失败处理和可衡量的输出质量。缺少这些，更像演示而不是可靠员工。
多 Agent 系统一定更好吗？
不一定。多 Agent 适合角色分工或并行任务，但会增加协调成本。对很多重复业务流程来说，单任务 Agent 或工作流 Agent 反而更稳定。
为什么 Agent 页面必须写安全细节？
因为 Agent 会执行动作。只要涉及文件、API、客户数据、付款或发布系统，就应说明权限、隔离边界、审批步骤和日志能力。