AI engineering comparisons
Tools in this field change every quarter, so each comparison here states when it was checked, compares on the criteria that decide real projects, and ends with when to pick which. The best tool for your project is the one that passes your own evals at a cost and operational load your team can carry, which a comparison page can narrow down but can't decide for you.
How to use these pages
1 - Start from your problem and constraints: data location, scale, team skills, budget. 2 - Use the comparison to shortlist two options. 3 - Build the same small slice with both, on your data. 4 - Compare on your evals, latency, cost and how easy it was to debug.
Comparisons
| Comparison | Decides |
|---|---|
| LangGraph vs CrewAI vs Microsoft Agent Framework | Agent orchestration |
| pgvector vs Qdrant vs Pinecone | Vector storage |
| LLM API providers | Which hosted models |
| vLLM vs SGLang vs Ollama | Serving open models |
| LlamaIndex vs LangChain | RAG and app frameworks |
| RAG vs fine-tuning | Knowledge vs behaviour |
| MCP vs function calling | Tool integration |
| Hosted vs self-hosted LLMs | Where models run |
Related choices covered elsewhere: long context vs retrieval and workflows vs agents.
Plain code is always an option
Frameworks save time on common patterns and add abstractions you'll need to debug through. For one agent with two tools, a provider SDK and your own loop is often clearer, and you can adopt a framework when state, approvals or recovery become real requirements.
From the floor
In AI Engineer HQ we use different tools for different purposes: LangGraph for the DevFix agent, where we needed stateful loops, sandboxing and test feedback, and CrewAI in a workshop where the goal was to build a multi-agent system and understand roles quickly.