I want to move beyond just using GPT-4 in a web browser. I’m looking to code a multi-agent system from scratch. Which libraries are currently the industry standard for AI Agents & Automation? I’ve heard about AutoGPT, but is it stable enough for production, or should I stick to something like CrewAI or Microsoft’s Semantic Kernel?
3 answers
If you are looking for production stability, I would steer clear of the early "experimental" autonomous agents like AutoGPT for now—they tend to get stuck in infinite loops. CrewAI is currently a favorite because it emphasizes "role-playing" where you can define a Manager agent and Worker agents with specific tasks. Another powerhouse is LangGraph, which gives you much finer control over the cyclical nature of agentic reasoning. For enterprise-grade integration, Semantic Kernel is excellent if you are already in the C# or Azure ecosystem, but for pure Python flexibility, CrewAI plus LangChain is the way to go.
Is it possible to run these agents locally using Llama 3 to save on API costs?
Don't forget about Haystack. It's very underrated for RAG-based agent workflows and has a very clean API.
Good point, Melissa. Haystack’s pipeline component makes it very easy to swap out different LLM providers as your project grows.
Absolutely, Patrick! Using Ollama or vLLM to host Llama 3 locally is a great way to build. Just keep in mind that "agentic" tasks require strong reasoning, so the 70B model usually performs significantly better than the 8B model when it comes to following complex multi-step instructions without losing the plot.