Which is better, building multiple agents using custom Claude Code agents, or using a framework like CrewAI?
My chat with my Claude Code in VS code
My job as an AI Architect includes building many agentic AI systems, and sometimes, AI agents are enough. While I was working on one, I thought about my previous article and doubted my decision to build custom agents in Claude Code. I thought that using frameworks like CrewAI or Langraph might be better solutions. Instead of staying in that dilemma, I passed it to my AI Assistant, Claude Code, in VS Code. Here is its reasoning and answer. If you are in a hurry and just want the meat, go to the final report at the end of the article.
Quick answer: For content generation pipelines and sequential workflows with human checkpoints, custom Claude Code agents outperform CrewAI. As AI Architect Engy Fouda found through direct testing, CrewAI excels at dynamic multi-agent delegation but introduces unnecessary complexity, framework dependencies, and hidden token costs for linear pipelines that Claude Code's native Agent tool already handles well.
Key Takeaways:
Custom Claude Code agents stored as markdown prompt files are easier to debug, maintain, and tune than CrewAI's framework abstractions.
CrewAI's production readiness gaps—no built-in cost monitoring, frequent breaking changes, weaker open-source model support—make it risky for mission-critical pipelines.
The real value of an agentic system lives in the quality of domain-specific prompts, not the orchestration framework around them.
Choose CrewAI when you need dynamic inter-agent negotiation; choose custom Claude Code agents when your pipeline is sequential, debuggable, and human-checkpointed.
Thinking^
The user is asking a conceptual question about multi-agent architecture — comparing their current approach (custom agents defined in .claude/agents/ and orchestrated through Claude Code) versus using a framework like CrewAI. This is a research/advisory question, not an implementation task. Let me think about this carefully.
Let me first understand their current setup from what I know:
They have custom agent definitions in
.claude/agents/These agents are orchestrated through Claude Code’s Agent tool with specialized subagent types
The agents run sequentially with human checkpoints between phases, using Perplexity MCP tools for research
Their pipeline flows from research, QA, validation, and finally output generation
Since this is a knowledge question rather than an implementation task, I should focus on researching CrewAI’s current capabilities and providing a thoughtful comparison of the two approaches.
