What should I use to build AI agents?

It depends on what kind of “AI agent” you want to build, but here’s the practical shortlist:

Best options by use case

1. Simple agent / workflow automation

• LangChain or LlamaIndex
• Good for: tool use, retrieval, chatbots, basic agent loops
• Best if you want to move fast and don’t need heavy orchestration

2. Multi-agent systems / coordination

• AutoGen (Microsoft)
• CrewAI
• Good for: multiple specialized agents, task delegation, collaboration
• Best if you want agents that talk to each other or split work

3. Production-grade agent workflows

• Temporal, Prefect, or Dagster combined with an LLM framework
• Good for: reliability, retries, state, long-running workflows
• Best if the agent needs to be robust in production

4. RAG-heavy agents

• LlamaIndex
• Good for: document retrieval, knowledge bases, grounded responses
• Best if the agent needs to answer from your data

5. Custom / low-level control

• Direct use of the OpenAI API or other model APIs
• Good for: full control, simpler architectures, less framework lock-in
• Best if you can design your own agent loop

---

My general recommendation

• Prototype fast: use LangChain or LlamaIndex
• Multi-agent collaboration: use AutoGen or CrewAI
• Production reliability: use Temporal + an LLM API
• Want maximum control: build directly on the model API

---

Important note

Frameworks help, but the core of a good agent is usually:

• a clear goal
• tool access
• memory/state
• retrieval over relevant data
• guardrails and evaluation

If you tell me:

1. what the agent should do,
2. your preferred language, and
3. whether this is a prototype or production system,

I can recommend the best stack specifically for you.