CrewAI幾正!——我嘅First Impressions
用CrewAI整AI-generated podcasts 10日後,我學到雖然呢個framework remarkably容易用,高質內容仲係需要真正嘅人類effort。
我幾星期前透過課程「Practical Multi AI Agents and Advanced Use Cases with crewAI」偶然發現CrewAI。雖然我一開始有啲hesitant去試,因為我已經用緊Langgraph做我嘅agent Sydney,但有兩件事spark咗我嘅興趣:
- 我一直喺explore點由scratch create podcasts,特別係點用Gen AI嚟「reason」over我15年嘅blog內容而唔係just copy and paste posts。我試過用NotebookLLM做呢個(效果okay,你可以聽podcast episode 呢度),但我想要更多control over conversation flow同opening hooks。
- 我想學啲新嘢同experiment OpenAI嘅text-to-speech models。
2025年11月更新:我已經graduated from用agent framework,開始用Claude Code同Google Gemini CLI由頭寫code。呢個俾我多好多control。我最新launch嘅係STRATUM(market intelligence)同DIALOGUE(AI podcast generator)——直接由下面嘅CrewAI experiment演變而嚟。
玩咗CrewAI大約10日之後,以下係我嘅key observations:
1. 正嘢
1.1 出奇咁容易上手
- 初始learning curve remarkably短——我幾個鐘就up and running
- Set up custom tools(好似用Weaviate作為vector store retrieve我blog嘅content)好straightforward。將podcast script轉成audio output都幾easy to set up。
- 用YAML files以plain English describe agents同佢哋嘅tasks嘅能力好powerful(Pro tip:Visual Studio Code嘅autocomplete喺呢度超helpful!)
1.2 靈活嘅model selection
Switch唔同LLMs簡單到只需要update你嘅crew.py:
llm_openai_4o_mini = LLM(model="gpt-4o-mini", temperature=0)
llm_anthropic_35 = LLM(model="claude-3-5-sonnet-20240620", temperature=0)
llm_openai_4o = LLM(model="gpt-4o", temperature=0)
llm_gemini_15_pro = LLM(model="gemini/gemini-1.5-pro-002", temperature=0)
你可以根據佢哋嘅strengths assign specific models俾唔同agents。例如,
@agent
def content_researcher(self) -> Agent:
return Agent(
config=self.agents_config['content_researcher'],
llm=llm_anthropic_35,
tools=[BlogContentRetrievalTool()],
verbose=True
)
呢一切意味住我對podcast嘅structure同scripting有great deal of control。
1.3 Text-to-Speech:Promise同Limitations
雖然OpenAI嘅text-to-speech API喺quality上impressive,但佢目前只提供六個voice models。對podcast creation嚟講,呢個幾limiting——特別係當你想create engaging嘅多host對話。Voice variety嘅缺乏意味住你嘅podcast可能聽落同其他用同樣technology嘅podcast類似。呢個definitely係一個我希望未來見到improvement嘅area,無論係OpenAI expand佢哋嘅voice options,定係透過integration with其他text-to-speech providers。
我都理解AI safety concerns,所以唔同嘅AI labs可能唔會太急住提供太多voice models。
2. 現實check:唔係「Click and Create」咁簡單
一開始,我擔心呢個可能contribute to我哋online見到嘅AI-generated content(或者「AI Slop」)嘅flood。畢竟,我可以喺大約5分鐘generate一個15分鐘嘅podcast script。(喺上面嘅DeepLearning.AI training course入面,Joao actually行咗一個「Content creation at scale」嘅code example。)
但我嘅perspective喺actually讀完/review最初幾個generate嘅scripts之後改變咗。
Create高質內容仲係需要significant work!
2.1 Thoughtful Agent結構
我要revise AI crew嘅結構好多次,加入additional roles,特別係「fact_checker」嘅角色。我而家嘅podcast crew包括:
- Content researcher
- Script writer
- Fact checker
- Script editor
- Audio producer
2.2 持續Refinement
成功需要:
- 用行業特定語言carefully define每個agent嘅goals同tasks。有actual行業podcast經驗嘅人可以用行業特定語言同ask每個agent perform非常specific嘅task。Output會好好多。
- Selective咁俾tool access(more唔一定better)。Agents好容易stuck喺continuous loops。
- Clear嘅agents之間delegation rules
- Specific嘅output structure requirements
- Well-defined嘅「What good work looks like」quality criteria(我仲拉埋我個女入嚟——佢係creative嗰個!:D)。
所以again如你所見,雖然用一crew AI agents幫我significantly加速工作(由research到scripting、fact check、revision、audio creation至少5X),最終都係要靠我create高質內容。
2.3 Model selection matters
唔同LLM models有distinct嘅「personalities」同varying levels嘅instruction-following。所以你要experiment嚟understand唔同models嘅strengths同weaknesses,同佢哋點suit你每個步驟嘅need。
一啲observations:
- 同一個model嘅API responses可以同web chat interface responses唔同
- 目前,用API做long-form content我prefer Anthropic models。但web version嚟講,我actually覺得claude-3-5-sonnet-20241022同GPT-4o on par。
- OpenAI嘅o1-preview係我做coding tasks嘅go-to
2.4 Feedback同Memory係Game-Changers
你一定要provide feedback俾你嘅AI crew。佢哋擅長follow directions但佢哋唔知你想要乜,亦read唔到你嘅mind(至少暫時未得 haha)。Train你crew through feedback嘅能力係crucial。
用CrewAI,train你嘅crew同give feedback好simple,只需要run:
crewai train -n <n_iterations> <filename> (optional)
雖然我未fully explore CrewAI嘅memory functions,但feedback同memory嘅combination對create consistent、高質output seems incredibly powerful。
3. Show Me The Results!
Ok, ok——我聽到你講「Chandler,你講夠喇。Show我一個你AI crew generate嘅podcast script sample!」
以下係complete workflow example:
- Research Phase:睇Content Researcher agent點analyze同extract key information from我嘅blog posts
- Fact Check Summary:Fact Checker嘅detailed verification report
- Initial Script Draft:Script Writer嘅first take on podcast conversation
- Final Polished Script:Script Editor refined嘅版本,improved flow同engagement
- Listen to the Result:Audio Producer agent produce嘅final audio版本
以上每個link show咗由raw content到polished podcast嘅progression,demonstrate唔同agents點contribute到最終product。
雖然我仲有thoughts關於點improve成個pipeline,我希望以上俾到你一個good sense of what is possible。
最後感想
CrewAI以佢嘅simplicity同power嘅balance impress咗我。雖然佢令content creation更accessible,但佢唔係magic button——quality仲係需要expertise、careful planning同continuous refinement。
你有冇experiment過multi-agent frameworks好似CrewAI或Langgraph?我好想聽你用佢哋build緊乜——feel free留言或reach out。
祝好,
Chandler
2026年1月更新:呢個CrewAI experiment最終成為咗DIALOGUE——一個full production app。我上面提到嘅voice limitation?我最後轉咗去Gemini TTS,俾我30個voices across 7種語言。同6個OpenAI voices比真係quite the upgrade!如果你curious嘅話,睇吓。





