SmolAgents

SmolAgents is a breath of fresh air if you're tired of sprawling agent frameworks that obscure the logic behind layers of abstractions. The core philosophy is 'agents that think in code' — your agent writes Python snippets as its actions, rather than emitting JSON tool calls. That means you get the full expressiveness of a programming language for complex decision-making, and the execution traces are literally code you can read. The ~1,000-line core is a masterclass in minimalism, and it's genuinely possible to read the whole `agents.py` in an afternoon. The model-agnostic support is a standout: you can swap between a local Transformers model, an Ollama instance, any provider on the Hugging Face Hub via `InferenceClientModel`, or any of 100+ LLMs through LiteLLM. You can also point at OpenAI-compatible endpoints like Together AI or OpenRouter, plus native integrations for Anthropic, Azure OpenAI, and Amazon Bedrock. This flexibility is a huge win if you're building a portfolio of agents that need to run across different backends. Modality-agnostic input — text, vision, video, audio — is another forward-looking feature. The tutorial on vision shows how to feed images to an agent, and the underlying support for video and audio inputs means you can build multimodal agents without a separate pipeline. The weaknesses are equally clear. There's no built-in durable state, checkpointing, or memory layer — you have to bolt that on yourself. Multi-agent orchestration is not a first-class citizen; if you need intricate task dependencies, you'll end up hand-rolling coordination. And while you can use tools from MCP, LangChain, or Hub Spaces, the pre-built tool ecosystem is smaller than LangChain's. Security is a real consideration: because the agent executes arbitrary code, running it on your local machine is risky. The library supports sandboxed execution via Blaxel, E2B, Modal, or Docker, and you should use one when running untrusted tasks. But even with sandboxing, there's overhead to setting up and managing those environments. In short, SmolAgents is not a production platform. It's a sharp tool for prototyping, research, and education. If you want to understand agent architectures from first principles, or you need to quickly test a code-driven agent idea against various models, it's an excellent choice. If you need a turnkey agent with memory, orchestration, and UI, look elsewhere — LangChain, CrewAI, or AutoGen will save you time.