Transformers Code Agent beats GAIA benchmark; migrate from transformers.agents to smolagents

AI Impact Summary

Transformers' Code Agent achieved top GAIA benchmark results, showing an agent-based approach with code actions can outperform traditional multi-step policies. The team migrated from transformers.agents to the stand-alone smolagents library, enabling a CodeAgent implementation that relies on web browsing, file inspection, and a safe Python interpreter to solve tough, tool-using tasks. They note code-based actions can be 30% fewer steps and cheaper in tokens than JSON, implying potential cost and latency benefits in production. This demonstrates a concrete migration path to a modern agent framework and provides practical patterns for building robust agentic systems.