AI Agent 学习笔记 – CrewAI

Agents: 介绍

Thu, 13 Nov 2025 00:00:00 +0000

https://www.crewai.com/

Streamline workflows across industries with powerful AI agents. Build and deploy automated workflows using any LLM and cloud platform.

使用强大的 AI 智能体简化各行业工作流程。使用任何 LLM 和云平台构建和部署自动化工作流程。

介绍

https://docs.crewai.com/en/introduction

Build AI agent teams that work together to tackle complex tasks

构建协同工作以处理复杂任务的 AI 代理团队

CrewAI 是一个轻量级、极速的 Python 框架，完全从头构建—完全独立于 LangChain 或其他智能体框架。

CrewAI 为开发者提供高级简洁性和精确的低级控制能力，非常适合创建针对任何场景量身定制的自主 AI 代理：

CrewAI 团队：优化自主性和协作智能，使您能够创建 AI 团队，其中每个代理都有特定的角色、工具和目标。
CrewAI Flows：实现细粒度、事件驱动的控制，通过单次 LLM 调用进行精确的任务编排，并原生支持 Crews。

CrewAI 框架

组件	描述	关键特性
Crew	顶级组织	• 管理 AI 代理团队 • 监督工作流程 • 确保协作 • 交付成果
AI Agents		• 担任特定角色（研究员、作者） • 使用指定的工具 • 可以委派任务 • 做出自主决策
Process	工作流管理系统	• 定义协作模式 • 控制任务分配 • 管理交互 • 确保高效执行
Tasks	单独任务	• 有明确目标 • 使用特定工具 • 输入到更大的流程中 • 产生可操作的结果

一切如何协同工作

Crew 组织整体运作
AI Agents 专注于他们的专业任务
Process 确保顺畅协作
Tasks 完成以实现目标

关键特性

基于角色的智能体

创建具有明确角色、专业知识和目标的专门化智能体 - 从研究员到分析师再到作家
灵活的工具

为代理配备自定义工具和 API，以与外部服务和数据源进行交互
智能协作

Agent 们协同工作，分享见解并协调任务以实现复杂目标
任务管理

定义顺序或并行工作流，让智能体自动处理任务依赖关系

工作流程如何运作

虽然团队擅长自主协作，但流程提供了结构化的自动化，对工作流执行提供细粒度控制。流程确保任务可靠、安全且高效地执行，精确处理条件逻辑、循环和动态状态管理。流程与团队无缝集成，使您能够在高度自主和精确控制之间取得平衡。

组件	描述	关键特性
Flow	结构化工作流编排	• 管理执行路径 • 处理状态转换 • 控制任务序列 • 确保可靠执行
Events	工作流操作的触发器	• 启动特定流程 • 实现动态响应 • 支持条件分支 • 允许实时调整
States	工作流执行上下文	• 维护执行数据 • 启用持久化 • 支持可恢复性 • 确保执行完整性
Crew Support	增强工作流程自动化	• 在需要时注入自主决策的空间 • 补充结构化工作流程 • 平衡自动化与智能 • 实现自适应决策

核心能力

事件驱动编排

定义精确的执行路径，对事件做出动态响应
精细控制

安全高效地管理工作流状态和条件执行
原生Crew 集成

轻松与团队结合，提升自主性和智能
确定性执行

通过显式的控制流程和错误处理确保可预测的结果

何时使用 Crews 与 Flows

理解何时使用 Crews 而非 Flows 是在您的应用程序中最大化 CrewAI 潜力的关键。

Use Case	推荐方法	Why? 为什么？
开放性研究	Crews	当任务需要创造性思维、探索和适应时
内容生成	Crews	用于协作创建文章、报告或营销材料
决策工作流	Flows	当您需要具有精确控制的、可审计的可预测决策路径时
API orchestration API 编排	Flows	与多个外部服务按特定顺序进行可靠集成
Hybrid applications 混合应用程序	综合方法	使用 Flow 编排整体流程，让 Crew 处理复杂的子任务

决策框架

在以下情况选择 Crew：您需要自主解决问题、创意协作或探索性任务
选择 Flows 的情况：当您需要确定性结果、可审计性或对执行的精确控制时
结合使用：当您的应用程序既需要结构化流程又需要自主智能的片段时，请结合两者

为什么选择 CrewAI？

🧠 自主操作：Agent 根据其角色和可用工具做出智能决策
📝 自然交互：Agent 像人类团队成员一样进行沟通和协作
🛠️ 可扩展设计：轻松添加新工具、角色和能力
🚀 生产就绪：为真实世界应用中的可靠性和可扩展性而构建
🔒 安全优先：专为满足企业安全需求而设计
💰 成本效益：优化以最小化 token 使用和 API 调用

Agents: 安装

Thu, 13 Nov 2025 00:00:00 +0000

https://docs.crewai.com/en/installation

准备工作

安装 uv

curl -LsSf https://astral.sh/uv/install.sh | sh

安装 CrewAI

uv tool install crewai

安装完成后验证：

uv tool list

可以看到输出:

crewai v1.6.1
- crewai

更新

如果需要更新版本，可以：

uv tool install crewai --upgrade

创建测试项目

mkdir -p ~/work/code/agents/crewai
cd ~/work/code/agents/crewai
crewai create crew testproject

按照提示选择：


Creating folder testproject...
Cache expired or not found. Fetching provider data from the web...
Downloading  [####################################]  961688/50589
Select a provider to set up:
1. openai
2. anthropic
3. gemini
4. nvidia_nim
5. groq
6. huggingface
7. ollama
8. watson
9. bedrock
10. azure
11. cerebras
12. sambanova
13. other
q. Quit
Enter the number of your choice or 'q' to quit: 1
Select a model to use for Openai:
1. gpt-4
2. gpt-4.1
3. gpt-4.1-mini-2025-04-14
4. gpt-4.1-nano-2025-04-14
5. gpt-4o
6. gpt-4o-mini
7. o1-mini
8. o1-preview
q. Quit
Enter the number of your choice or 'q' to quit: 2
Enter your OPENAI API key (press Enter to skip): 
API keys and model saved to .env file
Selected model: gpt-4.1
  - Created testproject/.gitignore
  - Created testproject/pyproject.toml
  - Created testproject/README.md
  - Created testproject/knowledge/user_preference.txt
  - Created testproject/src/testproject/__init__.py
  - Created testproject/src/testproject/main.py
  - Created testproject/src/testproject/crew.py
  - Created testproject/src/testproject/tools/custom_tool.py
  - Created testproject/src/testproject/tools/__init__.py
  - Created testproject/src/testproject/config/agents.yaml
  - Created testproject/src/testproject/config/tasks.yaml
Crew testproject created successfully!

实际项目结构：

.
└── testproject
    ├── knowledge
    │   └── user_preference.txt
    ├── pyproject.toml
    ├── README.md
    ├── src
    │   └── testproject
    │       ├── config
    │       │   ├── agents.yaml
    │       │   └── tasks.yaml
    │       ├── crew.py
    │       ├── __init__.py
    │       ├── main.py
    │       └── tools
    │           ├── custom_tool.py
    │           └── __init__.py
    └── tests

8 directories, 10 files

执行 install 命令来安装各种依赖：

crewai install

创建 .env 文件，内容为：

MODEL=openai/gpt-4.1
OPENAI_API_KEY=sk-or-v1-3d32348b8f97ab78a2510f0b60xxxxxxxxxxxxxxxxxxxxxxxxxdd2
OPENAI_API_BASE=https://openrouter.ai/api/v1

这里我用 openrouter 来替代 openai。

cd ~/work/code/agents/crewai/testproject
crewai run

输出非常的长：

Running the Crew
╭─────────────────────────────────────────────── Crew Execution Started ───────────────────────────────────────────────╮
│                                                                                                                      │
│  Crew Execution Started                                                                                              │
│  Name: crew                                                                                                          │
│  ID: 92af3cd2-638e-4e6a-b014-54b7aafc216c                                                                            │
│  Tool Args:                                                                                                          │
│                                                                                                                      │
│                                                                                                                      │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

🚀 Crew: crew
└── 📋 Task: research_task (ID: d0325f3f-b890-4bf5-b973-f216ff60e939)
    Status: Executing Task...
    └── 🧠 Thinking...
╭────────────────────────────────────────────────── 🤖 Agent Started ──────────────────────────────────────────────────╮
│                                                                                                                      │
│  Agent: AI LLMs Senior Data Researcher                                                                               │
│                                                                                                                      │
│  Task: Conduct a thorough research about AI LLMs Make sure you find any interesting and relevant information given   │
│  the current year is 2025.                                                                                           │
│                                                                                                                      │
│                                                                                                                      │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯


🚀 Crew: crew
└── 📋 Task: research_task (ID: d0325f3f-b890-4bf5-b973-f216ff60e939)
    Assigned to: AI LLMs Senior Data Researcher
    
    Status: ✅ Completed
╭─────────────────────────────────────────────── ✅ Agent Final Answer ────────────────────────────────────────────────╮
│                                                                                                                      │
│  Agent: AI LLMs Senior Data Researcher                                                                               │
│                                                                                                                      │
│  Final Answer:                                                                                                       │
│  - 1. **Multimodal LLMs Go Mainstream**: Large language models in 2025 now routinely integrate text, images, video,  │
│  audio, and even code, leading to “unified” models (like GPT-5, Gemini Ultra, and OpenAI's multimodal offerings)     │
│  capable of comprehending and generating across modalities without special prompts.                                  │
│                                                                                                                      │
│  - 2. **Agentic LLMs & Autonomous Reasoning**: Leading LLMs incorporate agentic capabilities for complex planning,   │
│  tool use, self-correction, and contextual adaptation, enabling them to accomplish multi-step tasks, act on web      │
│  data, and interact with external APIs safely and reliably.                                                          │
│                                                                                                                      │
│  - 3. **Open-Source Advancements**: Projects like Meta’s Llama 3 and Mistral’s Mixtral families have reached         │
│  competitive or superior performance compared to proprietary counterparts, democratizing access to cutting-edge      │
│  LLMs with permissive licensing—spurring grassroots innovation and wide deployment.                                  │
│                                                                                                                      │
│  - 4. **Fine-Tuning and Customization at Scale**: New approaches such as Differential Instruction Tuning, Reward     │
│  Modeling, and efficient Domain Adaptation allow organizations to rapidly specialize LLMs for legal, medical,        │
│  scientific, and enterprise applications with unprecedented accuracy and safety.                                     │
│                                                                                                                      │
│  - 5. **Ethics, Safety, and Alignment Breakthroughs**: As LLMs grow in complexity, novel alignment techniques        │
│  (Constitutional AI v2, multi-agent oversight, universal red-teaming) and robust watermarking/testing protocols      │
│  have substantially reduced toxic, biased, or hallucinated outputs, and improved explainability.                     │
│                                                                                                                      │
│  - 6. **Real-Time Reasoning & Edge Deployment**: Thanks to algorithmic and hardware advances (Transformers with      │
│  sparse attention, quantization, memory-efficient inference), powerful LLMs are now able to run on edge              │
│  devices—from smartphones to industrial IoT—enabling private, low-latency AI.                                        │
│                                                                                                                      │
│  - 7. **Ultra-Long Context Windows**: Models such as GPT-5 and Claude 3 Opus support context windows upwards of 1    │
│  million tokens, facilitating analysis and synthesis across entire books, massive codebases, and complex,            │
│  continuous conversations.                                                                                           │
│                                                                                                                      │
│  - 8. **AI-Powered Research & Discovery**: LLMs are now core assistants in science and engineering, accelerating     │
│  hypothesis generation, literature review, code synthesis, simulation, and experiment design in pharmaceuticals,     │
│  climate modeling, mathematics, and beyond.                                                                          │
│                                                                                                                      │
│  - 9. **Multilingual Capabilities and Global Expansion**: State-of-the-art LLMs boast fluent comprehension and       │
│  generation in 200+ languages, with dialect and context awareness, breaking down language barriers for education,    │
│  commerce, and international collaboration.                                                                          │
│                                                                                                                      │
│  - 10. **Personalization and Privacy**: Advances in federated learning, on-device fine-tuning, and                   │
│  privacy-preserving inference allow users to privately adapt LLMs to their personal style, preferences, and          │
│  knowledge—without ever sending sensitive data to the cloud.                                                         │
│                                                                                                                      │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

╭────────────────────────────────────────────────── Task Completion ───────────────────────────────────────────────────╮
│                                                                                                                      │
│  Task Completed                                                                                                      │
│  Name: research_task                                                                                                 │
│  Agent: AI LLMs Senior Data Researcher                                                                               │
│                                                                                                                      │
│  Tool Args:                                                                                                          │
│                                                                                                                      │
│                                                                                                                      │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

🚀 Crew: crew
├── 📋 Task: research_task (ID: d0325f3f-b890-4bf5-b973-f216ff60e939)
│   Assigned to: AI LLMs Senior Data Researcher
│   
│   Status: ✅ Completed
│   └── 🧠 Thinking...
└── 📋 Task: reporting_task (ID: aa65f26c-55ae-4d73-9cd6-4541403295c0)
    Status: Executing Task...
╭────────────────────────────────────────────────── 🤖 Agent Started ──────────────────────────────────────────────────╮
│                                                                                                                      │
│  Agent: AI LLMs Reporting Analyst                                                                                    │
│                                                                                                                      │
│  Task: Review the context you got and expand each topic into a full section for a report. Make sure the report is    │
│  detailed and contains any and all relevant information.                                                             │
│                                                                                                                      │
│                                                                                                                      │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯


🚀 Crew: crew
├── 📋 Task: research_task (ID: d0325f3f-b890-4bf5-b973-f216ff60e939)
│   Assigned to: AI LLMs Senior Data Researcher
│   
│   Status: ✅ Completed
└── 📋 Task: reporting_task (ID: aa65f26c-55ae-4d73-9cd6-4541403295c0)
    Status: Executing Task...
╭────────────────────────────────────────────────────── ✅ Agent Final Answer ──────────────────────────────────────────────────────╮
│                                                                                                                                   │
│  Agent: AI LLMs Reporting Analyst                                                                                                 │
│                                                                                                                                   │
│  Final Answer:                                                                                                                    │
│  # 2025 AI LLM Landscape: In-Depth Report                                                                                         │
│                                                                                                                                   │
│  ## 1. Multimodal LLMs Go Mainstream                                                                                              │
│                                                                                                                                   │
│  The year 2025 marks the full mainstream adoption of multimodal Large Language Models (LLMs), which now natively integrate text,  │
│  images, video, audio, and code within a unified architecture. Pioneering models such as GPT-5, Gemini Ultra, and the latest      │
│  offerings from OpenAI and Anthropic represent a transformative leap from text-only systems to truly universal models. These      │
│  models can seamlessly interpret, combine, and generate across diverse data types without requiring users to issue special        │
│  prompts or indicate the mode of information. For example, a physician can upload a radiology image, an audio dictation, and a    │
│  written case file, allowing the LLM to synthesize findings into a unified diagnostic report. In creative work, a user might ask  │
│  for a narrated storyboard with corresponding visuals and audio cues, all within a single conversational thread. This             │
│  convergence has made LLMs indispensable for multimedia content creation, data analysis, accessibility solutions, and real-world  │
│  task execution across industries.                                                                                                │
│                                                                                                                                   │
│  ## 2. Agentic LLMs & Autonomous Reasoning                                                                                        │
│                                                                                                                                   │
│  Modern LLMs are increasingly “agentic”—capable of autonomous reasoning, long-range planning, and dynamic interaction with        │
│  digital environments. Agentic LLMs can now devise multi-step plans, self-assess their actions, and execute complex workflows by  │
│  leveraging external APIs, web search, databases, and software tools. Key advances include toolformer models with integrated      │
│  plug-ins, self-correction and fallback mechanisms, as well as risk-based monitoring for operational safety. For instance, these  │
│  agentic models can automatically draft and send personalized email campaigns, analyze incoming responses, adapt messaging        │
│  strategies, and escalate exceptions—all with minimal human intervention. Safety protocols ensure that agentic LLMs reliably      │
│  handle web data, respect privacy boundaries, and maintain audit trails, making them suitable for high-stakes applications in     │
│  research, enterprise, and customer engagement.                                                                                   │
│                                                                                                                                   │
│  ## 3. Open-Source Advancements                                                                                                   │
│                                                                                                                                   │
│  Open-source LLMs have evolved rapidly, with Meta’s Llama 3 and Mistral’s Mixtral families delivering performance that matches    │
│  or exceeds major proprietary models on benchmarks of accuracy, efficiency, and adaptability. These models are released under     │
│  permissive licenses, fueling an explosion of grassroots contributions, community-driven research, and corporate adoption.        │
│  Open-source LLMs enable broader experimentation with model architectures, transparency, and custom deployment, lowering          │
│  barriers to innovation. This democratization of technology supports local language support, domain-specific tuning, and reduced  │
│  cost for startups and non-profit organizations. In 2025, government agencies, healthcare providers, and independent developers   │
│  increasingly rely upon open LLMs to craft solutions tailored to regional, ethical, and regulatory requirements.                  │
│                                                                                                                                   │
│  ## 4. Fine-Tuning and Customization at Scale                                                                                     │
│                                                                                                                                   │
│  The latest techniques in LLM specialization enable rapid, scalable, and secure adaptation for highly specific workflows or       │
│  domains. Differential Instruction Tuning allows LLMs to adjust to nuanced user instructions or evolving best practices without   │
│  full retraining. Reward Modeling integrates reinforcement learning from human or expert feedback to align LLM responses with     │
│  organizational values or compliance needs. Domain Adaptation leverages small, curated corpora or expert-in-the-loop corrections  │
│  for precision in fields such as finance, law, and medicine. These fine-tuning methods unlock state-of-the-art performance on     │
│  tailored tasks—legal contract drafting, scientific literature review, or technical troubleshooting—while preserving safety and   │
│  general language capabilities. Combined with improved data privacy protocols, this enables widespread enterprise and             │
│  vertical-market deployment.                                                                                                      │
│                                                                                                                                   │
│  ## 5. Ethics, Safety, and Alignment Breakthroughs                                                                                │
│                                                                                                                                   │
│  As LLMs increase in sophistication and reach, significant breakthroughs have occurred in ensuring ethical operation,             │
│  robustness, and user trust. Enhanced alignment strategies—such as Constitutional AI v2, which encodes explicit ethical           │
│  frameworks, and multi-agent oversight systems—provide continual safety checks and adversarial testing. Universal red-teaming     │
│  networks constantly probe for undesirable outputs. Built-in watermarking and robust traceability mechanisms attest to the        │
│  provenance and integrity of generated content. Improved explainability methods offer transparency into model decision-making,    │
│  facilitating responsible integration into regulated environments. As a result, rates of toxic, biased, or hallucinated outputs   │
│  have dropped precipitously, while accountability and compliance have risen, paving the way for safe, large-scale deployment.     │
│                                                                                                                                   │
│  ## 6. Real-Time Reasoning & Edge Deployment                                                                                      │
│                                                                                                                                   │
│  Advancements in both algorithms (e.g., sparse Transformers, quantization, retrieval-augmented generation) and hardware           │
│  (optimized AI accelerators) have enabled powerful LLMs to operate efficiently on edge devices. Smartphones, medical              │
│  instruments, autonomous machinery, and industrial IoT systems now run real-time LLM inference, preserving privacy, lowering      │
│  latency, and ensuring continuous operation even without cloud connectivity. For example, edge-deployed LLMs power intelligent    │
│  voice assistants, in-car navigation and diagnostics, and privacy-preserving wearable health monitors. These solutions are        │
│  resilient, highly available, and adapt to bandwidth or regulatory constraints, decentralizing AI and supporting new markets and  │
│  use cases.                                                                                                                       │
│                                                                                                                                   │
│  ## 7. Ultra-Long Context Windows                                                                                                 │
│                                                                                                                                   │
│  The expansion of LLM context windows to one million tokens (as in GPT-5 and Claude 3 Opus) has revolutionized how machines       │
│  process and connect information. These ultra-long context capabilities allow for the seamless ingestion and synthesis of entire  │
│  books, massive technical documentation, code repositories, full legal contracts, or multi-hour transcripts within a single       │
│  session. This fundamentally enhances long-form reasoning, document understanding, and persistent conversational memory. Use      │
│  cases include end-to-end review of clinical trials, comprehensive due diligence in finance, and lifelike, multi-session          │
│  personal assistants. The technical advances rely on memory-efficient inference and context-aware retrieval techniques,           │
│  maintaining high relevance, accuracy, and response times despite the scale.                                                      │
│                                                                                                                                   │
│  ## 8. AI-Powered Research & Discovery                                                                                            │
│                                                                                                                                   │
│  Modern LLMs have become essential assistants in scientific and engineering research, accelerating nearly every phase of the      │
│  discovery process. They autonomously generate hypotheses, aggregate and synthesize literature, design and simulate experiments,  │
│  analyze datasets, and draft publications. In pharmaceuticals, LLMs help identify new molecular structures, predict outcomes of   │
│  drug trials, and suggest modifications for increased efficacy. In climate science, they aid in modeling, scenario analysis, and  │
│  policy simulation. In mathematics and engineering, LLMs generate proofs, suggest optimization strategies, and support complex    │
│  design thinking. This synergy has shortened the cycle from research question to actionable insight, fueling rapid progress       │
│  across disciplines.                                                                                                              │
│                                                                                                                                   │
│  ## 9. Multilingual Capabilities and Global Expansion                                                                             │
│                                                                                                                                   │
│  State-of-the-art LLMs now deliver real-time, nuanced comprehension and generation in over 200 languages—covering regional        │
│  dialects, technical jargon, and cultural context. This profound multilingual ability breaks down language barriers for global    │
│  collaboration in education, commerce, healthcare, and government. Automated translation and summarization platforms,             │
│  globally-aware customer service agents, and context-sensitive educational tools empower individuals and businesses to            │
│  communicate fluidly across borders. For international partnerships and remote working environments, the LLM’s ability to adapt   │
│  to local context and etiquette ensures effective, respectful, and precise interactions.                                          │
│                                                                                                                                   │
│  ## 10. Personalization and Privacy                                                                                               │
│                                                                                                                                   │
│  The interplay of federated learning, on-device training, and privacy-preserving inference empowers users to maintain control     │
│  and confidentiality over their data while harnessing LLM personalization. Individuals and organizations can now tailor LLMs to   │
│  recognize specific terminologies, communication styles, interests, and proprietary knowledge, with all updates occurring         │
│  locally. Sensitive data—such as medical histories, financial information, and personal communications—does not leave the user’s  │
│  device, ensuring compliance with strict privacy regulations (e.g., GDPR, HIPAA). This private adaptation unlocks deeply          │
│  personalized AI-driven experiences: custom writing assistants, domain-specific advisors, and user-aware automation across        │
│  personal and professional domains.                                                                                               │
│                                                                                                                                   │
│  ---                                                                                                                              │
│                                                                                                                                   │
│  ## Conclusion                                                                                                                    │
│                                                                                                                                   │
│  The convergence of these ten trends has catapulted LLMs from experimental systems to omnipresent, transformative technologies    │
│  in 2025. Multimodality, agentic autonomy, open innovation, and sustained progress in safety and explainability underpin a new    │
│  era of AI impact. Widespread deployment across devices and languages, paired with scalable personalization and privacy,          │
│  broadens access and potential while mitigating risks. Organizations that integrate these advances will unlock unprecedented      │
│  productivity, insight, and global reach. The ongoing evolution of the AI LLM ecosystem promises to reshape industries, empower   │
│  individuals, and redefine how knowledge and intelligence are applied in the digital age.                                         │
│                                                                                                                                   │
╰───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
🚀 Crew: crew
├── 📋 Task: research_task (ID: d0325f3f-b890-4bf5-b973-f216ff60e939)
│   Assigned to: AI LLMs Senior Data Researcher
│   
│   Status: ✅ Completed
└── 📋 Task: reporting_task (ID: aa65f26c-55ae-4d73-9cd6-4541403295c0)
    Assigned to: AI LLMs Reporting Analyst
    
    Status: ✅ Completed
╭───────────────────────────────────────────────────────── Task Completion ─────────────────────────────────────────────────────────╮
│                                                                                                                                   │
│  Task Completed                                                                                                                   │
│  Name: reporting_task                                                                                                             │
│  Agent: AI LLMs Reporting Analyst                                                                                                 │
│                                                                                                                                   │
│  Tool Args:                                                                                                                       │
│                                                                                                                                   │
│                                                                                                                                   │
╰───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

╭───────────────────────────────────────────────────────── Crew Completion ─────────────────────────────────────────────────────────╮
│                                                                                                                                   │
│  Crew Execution Completed                                                                                                         │
│  Name: crew                                                                                                                       │
│  ID: 92af3cd2-638e-4e6a-b014-54b7aafc216c                                                                                         │
│  Tool Args:                                                                                                                       │
│  Final Output: # 2025 AI LLM Landscape: In-Depth Report                                                                           │
│                                                                                                                                   │
│  ## 1. Multimodal LLMs Go Mainstream                                                                                              │
│                                                                                                                                   │
│  The year 2025 marks the full mainstream adoption of multimodal Large Language Models (LLMs), which now natively integrate text,  │
│  images, video, audio, and code within a unified architecture. Pioneering models such as GPT-5, Gemini Ultra, and the latest      │
│  offerings from OpenAI and Anthropic represent a transformative leap from text-only systems to truly universal models. These      │
│  models can seamlessly interpret, combine, and generate across diverse data types without requiring users to issue special        │
│  prompts or indicate the mode of information. For example, a physician can upload a radiology image, an audio dictation, and a    │
│  written case file, allowing the LLM to synthesize findings into a unified diagnostic report. In creative work, a user might ask  │
│  for a narrated storyboard with corresponding visuals and audio cues, all within a single conversational thread. This             │
│  convergence has made LLMs indispensable for multimedia content creation, data analysis, accessibility solutions, and real-world  │
│  task execution across industries.                                                                                                │
│                                                                                                                                   │
│  ## 2. Agentic LLMs & Autonomous Reasoning                                                                                        │
│                                                                                                                                   │
│  Modern LLMs are increasingly “agentic”—capable of autonomous reasoning, long-range planning, and dynamic interaction with        │
│  digital environments. Agentic LLMs can now devise multi-step plans, self-assess their actions, and execute complex workflows by  │
│  leveraging external APIs, web search, databases, and software tools. Key advances include toolformer models with integrated      │
│  plug-ins, self-correction and fallback mechanisms, as well as risk-based monitoring for operational safety. For instance, these  │
│  agentic models can automatically draft and send personalized email campaigns, analyze incoming responses, adapt messaging        │
│  strategies, and escalate exceptions—all with minimal human intervention. Safety protocols ensure that agentic LLMs reliably      │
│  handle web data, respect privacy boundaries, and maintain audit trails, making them suitable for high-stakes applications in     │
│  research, enterprise, and customer engagement.                                                                                   │
│                                                                                                                                   │
│  ## 3. Open-Source Advancements                                                                                                   │
│                                                                                                                                   │
│  Open-source LLMs have evolved rapidly, with Meta’s Llama 3 and Mistral’s Mixtral families delivering performance that matches    │
│  or exceeds major proprietary models on benchmarks of accuracy, efficiency, and adaptability. These models are released under     │
│  permissive licenses, fueling an explosion of grassroots contributions, community-driven research, and corporate adoption.        │
│  Open-source LLMs enable broader experimentation with model architectures, transparency, and custom deployment, lowering          │
│  barriers to innovation. This democratization of technology supports local language support, domain-specific tuning, and reduced  │
│  cost for startups and non-profit organizations. In 2025, government agencies, healthcare providers, and independent developers   │
│  increasingly rely upon open LLMs to craft solutions tailored to regional, ethical, and regulatory requirements.                  │
│                                                                                                                                   │
│  ## 4. Fine-Tuning and Customization at Scale                                                                                     │
│                                                                                                                                   │
│  The latest techniques in LLM specialization enable rapid, scalable, and secure adaptation for highly specific workflows or       │
│  domains. Differential Instruction Tuning allows LLMs to adjust to nuanced user instructions or evolving best practices without   │
│  full retraining. Reward Modeling integrates reinforcement learning from human or expert feedback to align LLM responses with     │
│  organizational values or compliance needs. Domain Adaptation leverages small, curated corpora or expert-in-the-loop corrections  │
│  for precision in fields such as finance, law, and medicine. These fine-tuning methods unlock state-of-the-art performance on     │
│  tailored tasks—legal contract drafting, scientific literature review, or technical troubleshooting—while preserving safety and   │
│  general language capabilities. Combined with improved data privacy protocols, this enables widespread enterprise and             │
│  vertical-market deployment.                                                                                                      │
│                                                                                                                                   │
│  ## 5. Ethics, Safety, and Alignment Breakthroughs                                                                                │
│                                                                                                                                   │
│  As LLMs increase in sophistication and reach, significant breakthroughs have occurred in ensuring ethical operation,             │
│  robustness, and user trust. Enhanced alignment strategies—such as Constitutional AI v2, which encodes explicit ethical           │
│  frameworks, and multi-agent oversight systems—provide continual safety checks and adversarial testing. Universal red-teaming     │
│  networks constantly probe for undesirable outputs. Built-in watermarking and robust traceability mechanisms attest to the        │
│  provenance and integrity of generated content. Improved explainability methods offer transparency into model decision-making,    │
│  facilitating responsible integration into regulated environments. As a result, rates of toxic, biased, or hallucinated outputs   │
│  have dropped precipitously, while accountability and compliance have risen, paving the way for safe, large-scale deployment.     │
│                                                                                                                                   │
│  ## 6. Real-Time Reasoning & Edge Deployment                                                                                      │
│                                                                                                                                   │
│  Advancements in both algorithms (e.g., sparse Transformers, quantization, retrieval-augmented generation) and hardware           │
│  (optimized AI accelerators) have enabled powerful LLMs to operate efficiently on edge devices. Smartphones, medical              │
│  instruments, autonomous machinery, and industrial IoT systems now run real-time LLM inference, preserving privacy, lowering      │
│  latency, and ensuring continuous operation even without cloud connectivity. For example, edge-deployed LLMs power intelligent    │
│  voice assistants, in-car navigation and diagnostics, and privacy-preserving wearable health monitors. These solutions are        │
│  resilient, highly available, and adapt to bandwidth or regulatory constraints, decentralizing AI and supporting new markets and  │
│  use cases.                                                                                                                       │
│                                                                                                                                   │
│  ## 7. Ultra-Long Context Windows                                                                                                 │
│                                                                                                                                   │
│  The expansion of LLM context windows to one million tokens (as in GPT-5 and Claude 3 Opus) has revolutionized how machines       │
│  process and connect information. These ultra-long context capabilities allow for the seamless ingestion and synthesis of entire  │
│  books, massive technical documentation, code repositories, full legal contracts, or multi-hour transcripts within a single       │
│  session. This fundamentally enhances long-form reasoning, document understanding, and persistent conversational memory. Use      │
│  cases include end-to-end review of clinical trials, comprehensive due diligence in finance, and lifelike, multi-session          │
│  personal assistants. The technical advances rely on memory-efficient inference and context-aware retrieval techniques,           │
│  maintaining high relevance, accuracy, and response times despite the scale.                                                      │
│                                                                                                                                   │
│  ## 8. AI-Powered Research & Discovery                                                                                            │
│                                                                                                                                   │
│  Modern LLMs have become essential assistants in scientific and engineering research, accelerating nearly every phase of the      │
│  discovery process. They autonomously generate hypotheses, aggregate and synthesize literature, design and simulate experiments,  │
│  analyze datasets, and draft publications. In pharmaceuticals, LLMs help identify new molecular structures, predict outcomes of   │
│  drug trials, and suggest modifications for increased efficacy. In climate science, they aid in modeling, scenario analysis, and  │
│  policy simulation. In mathematics and engineering, LLMs generate proofs, suggest optimization strategies, and support complex    │
│  design thinking. This synergy has shortened the cycle from research question to actionable insight, fueling rapid progress       │
│  across disciplines.                                                                                                              │
│                                                                                                                                   │
│  ## 9. Multilingual Capabilities and Global Expansion                                                                             │
│                                                                                                                                   │
│  State-of-the-art LLMs now deliver real-time, nuanced comprehension and generation in over 200 languages—covering regional        │
│  dialects, technical jargon, and cultural context. This profound multilingual ability breaks down language barriers for global    │
│  collaboration in education, commerce, healthcare, and government. Automated translation and summarization platforms,             │
│  globally-aware customer service agents, and context-sensitive educational tools empower individuals and businesses to            │
│  communicate fluidly across borders. For international partnerships and remote working environments, the LLM’s ability to adapt   │
│  to local context and etiquette ensures effective, respectful, and precise interactions.                                          │
│                                                                                                                                   │
│  ## 10. Personalization and Privacy                                                                                               │
│                                                                                                                                   │
│  The interplay of federated learning, on-device training, and privacy-preserving inference empowers users to maintain control     │
│  and confidentiality over their data while harnessing LLM personalization. Individuals and organizations can now tailor LLMs to   │
│  recognize specific terminologies, communication styles, interests, and proprietary knowledge, with all updates occurring         │
│  locally. Sensitive data—such as medical histories, financial information, and personal communications—does not leave the user’s  │
│  device, ensuring compliance with strict privacy regulations (e.g., GDPR, HIPAA). This private adaptation unlocks deeply          │
│  personalized AI-driven experiences: custom writing assistants, domain-specific advisors, and user-aware automation across        │
│  personal and professional domains.                                                                                               │
│                                                                                                                                   │
│  ---                                                                                                                              │
│                                                                                                                                   │
│  ## Conclusion                                                                                                                    │
│                                                                                                                                   │
│  The convergence of these ten trends has catapulted LLMs from experimental systems to omnipresent, transformative technologies    │
│  in 2025. Multimodality, agentic autonomy, open innovation, and sustained progress in safety and explainability underpin a new    │
│  era of AI impact. Widespread deployment across devices and languages, paired with scalable personalization and privacy,          │
│  broadens access and potential while mitigating risks. Organizations that integrate these advances will unlock unprecedented      │
│  productivity, insight, and global reach. The ongoing evolution of the AI LLM ecosystem promises to reshape industries, empower   │
│  individuals, and redefine how knowledge and intelligence are applied in the digital age.                                         │
│                                                                                                                                   │
│                                                                                                                                   │
╰───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

╭───────────────────────────────────────────────────────── Tracing Status ──────────────────────────────────────────────────────────╮
│                                                                                                                                   │
│  Info: Tracing is disabled.                                                                                                       │
│                                                                                                                                   │
│  To enable tracing, do any one of these:                                                                                          │
│  • Set tracing=True in your Crew/Flow code                                                                                        │
│  • Set CREWAI_TRACING_ENABLED=true in your project's .env file                                                                    │
│  • Run: crewai traces enable                                                                                                      │
│                                                                                                                                   │
╰───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
➜

最后提示没有开启 tracing，参照提示，修改 .env 文件，加入一行:

CREWAI_TRACING_ENABLED=true

重新执行，这次会有 tracing 信息：

│ ✅ Trace batch finalized with session ID: 5122e089-0d27-4d3d-839e-bd39da8e44a5                                                    │
│                                                                                                                                   │
│ 🔗 View here:                                                                                                                     │
│ https://app.crewai.com/crewai_plus/ephemeral_trace_batches/5122e089-0d27-4d3d-839e-bd39da8e44a5?access_code=TRACE-1e03614a69      │
│ 🔑 Access Code: TRACE-1e03614a69

打开查看 tracing 信息，就两次 task，两次 LLM 调用：

Agents: 启动速度

Thu, 13 Nov 2025 00:00:00 +0000

部署

参考:

https://github.com/letta-ai/letta?tab=readme-ov-file#simple-hello-world-example

准备工作

以 python 为例:

pip install letta-client

export LETTA_API_KEY="sk-let-NTExNDRjZjxxxxxxx2OA=="

为了访问 letta 的服务器, 需要科学上网. 如果用的是 socks 代码, 则需要安装 pip 包:

pip install "httpx[socks]"

编写 agent

mkdir -p ~/work/code/agents/letta/hellowworld
cd ~/work/code/agents/letta/hellowworld

vi main.py

输入内容:

import os
import time
import psutil
from letta_client import Letta

# 记录程序开始时间
start_time = time.perf_counter()

# Connect to Letta Cloud (get your API key at https://app.letta.com/api-keys)
client = Letta(api_key=os.getenv("LETTA_API_KEY"))
# 如果是自托管，可以改成：
# client = Letta(base_url="http://localhost:8283", embedding="openai/text-embedding-3-small")

agent_state = client.agents.create(
    model="openai/gpt-4.1",
    memory_blocks=[
        {
          "label": "human",
          "value": "The human's name is Chad. They like vibe coding."
        },
        {
          "label": "persona",
          "value": "My name is Sam, a helpful assistant."
        }
    ],
    tools=["web_search", "run_code"]
)

# 在创建 agent 完成后，记录时间
elapsed = (time.perf_counter() - start_time) * 1000  # 毫秒

# 获取当前进程内存占用
process = psutil.Process()
mem = process.memory_info().rss / (1024 * 1024)  # MB

# 打印启动时间和内存占用
print(f"程序启动时间: {elapsed:.0f} ms, 内存占用: {mem:.2f} MB")

# 打印 agent id
print(agent_state.id)

response = client.agents.messages.create(
    agent_id=agent_state.id,
    messages=[
        {
            "role": "user",
            "content": "Hey, nice to meet you, my name is Brad."
        }
    ]
)

# 打印响应消息
for message in response.messages:
    print(message)

代码在官方 example 基础上增加了启动时间打印和内存使用打印.

运行:

# 开启 http 代理
# proxyon

python main.py

输出为:

程序启动时间: 1058 ms, 内存占用: 52.19 MB
agent-9d98c82e-4d4c-4d9e-a97d-187a30b52756
AssistantMessage(id='message-d567f53e-a4cc-42c2-9a24-4051c58fe305', content='Hey Brad, nice to meet you too! If there’s anything you want to work on or chat about, just let me know.', date=datetime.datetime(2025, 11, 21, 10, 35, 24, tzinfo=TzInfo(UTC)), is_err=None, message_type='assistant_message', name=None, otid='d567f53e-a4cc-42c2-9a24-4051c58fe300', run_id='run-3ae41158-73f2-45e7-be1b-3558254d9163', sender_id=None, seq_id=None, step_id='step-2535b417-86d8-4279-a56f-7e481f96ad72')

多测试几次, 数据为:

程序启动时间: 1052 ms, 内存占用: 52.30 MB
程序启动时间: 1173 ms, 内存占用: 52.66 MB
程序启动时间: 992 ms, 内存占用: 52.03 MB

所以启动时间大概是 1000 ms, 内存占用 52 MB.

注意: 这里启动不够快的主要原因可能是创建 client 时需要和服务器端建立连接, 而服务器端在国外访问速度慢, 如果服务器端在本地部署, 应该可以快很多.

注意2: 这里用快照启动没有意义, 因此 client 建立的连接在快照恢复时会无效, 只能冷启动.