smolagents: library to build agents that write actions in code

AI Impact Summary

smolagents introduces a minimal, code-centric agent framework that lets LLMs generate and execute actions in code through tools like DuckDuckGoSearchTool and models such as HfApiModel, with sandboxed execution via E2B. This lowers integration friction by enabling external-tool calls (Google Maps, weather APIs, knowledge bases) within agent workflows, allowing dynamic decision-making that previously required custom orchestration. It also positions the library as the successor to transformers.agents and enables cross-LLM workflows via Hub integrations and LiteLLM, potentially easing migration for teams currently using OpenAI or Anthropic APIs.