Hugging Face Smol Agents

💡 Hugging Face recently released Smol Agents.

‍

What sets Smol Agents apart is its singular focus: 𝘀𝗶𝗺𝗽𝗹𝗶𝗰𝗶𝘁𝘆.

‍

𝙃𝙤𝙬 𝙞𝙩 𝙬𝙤𝙧𝙠𝙨:

‍

Like many agent frameworks its based on the ReAct architecture (see below image), they support multi-agent systems through a manager agent, where agents can be nested inside each other, and manager agents can submit questions to their child agents.

‍

𝙒𝙝𝙖𝙩 𝙞𝙨 𝙝𝙖𝙨 𝙩𝙝𝙖𝙩 𝙤𝙩𝙝𝙚𝙧 𝙛𝙧𝙖𝙢𝙚𝙬𝙤𝙧𝙠𝙨 𝙙𝙤𝙣𝙩:

‍

𝗕𝗿𝗶𝗻𝗴 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝗟𝗟𝗠: You can use ANY LLM to power your agent as function calling happens by generating code, which is proven to be more effective than the Json that most agentic frameworks use.

‍

𝗦𝗺𝗼𝗼𝘁𝗵 𝗜𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻: with Hugging Face for Models & Open Telemetry (open source alternative to LangSmith) for logging of agent outputs. This logging is super important when diagnosing problems and further developing agents.

‍

𝗖𝗼𝗻𝗰𝗶𝘀𝗲 𝗮𝗻𝗱 𝗘𝘅𝗰𝗲𝗹𝗹𝗲𝗻t 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻: Some really useful nuggets on their page about agent development!

‍

https://huggingface.co/docs/smolagents/index

‍

𝙒𝙝𝙖𝙩 𝙞𝙩 𝙞𝙨 𝙢𝙞𝙨𝙨𝙞𝙣𝙜 𝙩𝙝𝙖𝙩 𝙤𝙩𝙝𝙚𝙧 𝙛𝙧𝙖𝙢𝙚𝙬𝙤𝙧𝙠𝙨 𝙝𝙖𝙫𝙚:

• Async Execution
• Memory
• Time Travel Debugging
• Multi-language Support

𝙒𝙝𝙚𝙧𝙚 𝙘𝙖𝙣 𝙄 𝙨𝙚𝙚 𝙞𝙩 𝙬𝙤𝙧𝙠𝙞𝙣𝙜 𝙬𝙚𝙡𝙡?

• Developing Proof of concept
• Research
• Testing suitability of open source models for Agent development
• Testing code execution tools locally

‍

𝙒𝙝𝙚𝙧𝙚 𝙘𝙖𝙣 𝙄 𝙣𝙤𝙩 𝙨𝙚𝙚 𝙞𝙩 𝙬𝙤𝙧𝙠𝙞𝙣𝙜 𝙬𝙚𝙡𝙡?

• Places where you want fine control - Like a customer facing conversational agent
• Any production setting where you need support for async execution, conversation memory, persistent state etc.

‍