Agent Framework Thoughts

• According to Anthropic blog post here it’s better to
  • Start simple, take time to analyse if agents are really needed.
  • Understand the difference between
    • (i) Workflow: augmented LLM (with retrieval, tools and memories) organized in pipeline with semi-rigid structure and loops to achive a set of tasks. The execution can be planned and behaviour is predictable
    • (ii) Agent: open-ended augmented LLM that define itself the structure of execution it needs to achieve his task.
• Same Mindset in Huggingface Blostpost about SmolAgents
  
  They produced a great table to understand the “agent spectrum”: how much of an agent you current LLM orchestration is.

There is a lot of Agent Framework out there. Anthropic suggest to not use any and instead go directly low level to implement your LLM calls. Here is a list of some framework I’ve seen and what I think about it:

• OpenAI Swarm: too much research oriented and linked to OpenAI
• PydanticAI: a lot of good feedback but not compatible yet with AWS Bedrock
• Smolagent: great but dependent on E2B for CodeAgent which is an absolute no go. (Paid sandboxed code execution environnement for agent). Can’t use it before they allow local solutions
• Langchain/Langraph: known to be a production nightmare, stay very far away.
• Haystack: What I use so far for RAG pipelines. But might be too over-engineered for easily create and modify agents… They provide a Cookbook however to replicate OpenAI Swarm ideas.