Hugging Face's smolagents makes code-writing AI agents dirt simple in 3 lines

Hugging Face just dropped smolagents, a new library that strips building AI agents down to its absolute bones. We’re talking three lines of Python to give a language model the power to search the web and reason through multi-step problems. The core idea is elegant and, frankly, long overdue: if an AI needs to act in the real world, let it write the code to do so directly rather than fumbling with rigid JSON function calls.

The library’s big bet is on “code agents.” Instead of the standard industry practice of having an LLM output a JSON blob like {"tool": "search", "query": "leopard speed"} which then gets parsed, the model writes executable Python snippets. The reasoning, as the Hugging Face team points out, is obvious when you think about it. We spent decades designing programming languages to express computer actions precisely. JSON schemas for tool use are a clumsy workaround. Research already shows code-based actions dramatically outperform JSON-based tool calling, and smolagents bakes this insight into a dead-simple API.

The article cuts through the agent hype with a refreshingly pragmatic framework. Agency isn’t a binary switch, but a spectrum. A hard-coded if/else workflow for customer requests has zero agency and is 100% reliable—if your use case fits. An agent should only enter the picture when predetermined workflows keep breaking, like a customer asking a tangled, multi-variable question about rescheduling a surf trip that requires checking weather, calendars, and maps. The library’s CodeAgent handles this by running a loop: the LLM proposes an action as code, the system executes it and feeds the observation back, repeating until the task is solved.

This launch feels like a natural step in the maturation of agentic systems. We’re moving past the era of complex orchestration frameworks that abstract away what the LLM is actually doing. By making the core loop transparent and putting code generation front and center, smolagents might just reset expectations for how developers build these things. The library is available now, and the team is promising more integrations and tools to come.