SmolAgents

Sweet spot: a developer or researcher who wants to build an agent and actually understand what is happening on every step. SmolAgents strips agents down to their essentials — an LLM, an action, a result, a new prompt — and that clarity is refreshing. It is also genuinely capable for a lot of real tasks: data analysis, research, web browsing with a code interpreter. Failure modes. For production multi-step agents with human-in-the-loop, durable state, and recovery semantics, SmolAgents does not compete with LangGraph — it was never designed to. Sandbox discipline is on you; running the CodeAgent with LocalPythonExecutor on untrusted input is a serious security risk. And the tool ecosystem, while growing, is thinner than the big frameworks. What to pilot. Pick a task that can be solved by writing and running code (a data-analysis question, a file-processing job, a research question that needs web fetches). Build a SmolAgents CodeAgent for it and compare to your current approach. If the code-writing action space makes the task easier, SmolAgents is the right tool; if the task is pure tool-calling with no computation, any framework will do.

SmolAgents is HuggingFace's intentionally tiny agent framework. The full source code is under 1,000 lines of Python, and the philosophy is explicitly anti-bloat: the primary loop is a CodeAgent that writes Python code to solve a task, executes it in a sandbox, and iterates. A ToolCallingAgent variant uses JSON function calls instead. The core insight: code-writing agents are more capable than tool-calling agents on complex reasoning tasks because code is a more expressive action space (variables, loops, conditionals) than a sequence of JSON function calls. SmolAgents bakes this philosophy in and keeps the scaffolding to the absolute minimum. It ships with integrations to HuggingFace Inference Endpoints, OpenAI, Anthropic, LiteLLM, and any OpenAI-compatible API. Tools can be pulled from the Hub or written as plain Python functions with type hints. Sandbox options include E2B, Docker, and a local Python executor (unsafe without sandboxing). Released by HuggingFace in late 2024 and quickly adopted as the "learn how agents work" framework — the entire codebase is readable in one sitting, which is rare in a space dominated by heavy abstractions.