Transformers Code Agent beats GAIA benchmark — migration to smolagents announced
AI Impact Summary
A Transformer-based Code Agent built with transformers.agents achieves top results on the GAIA benchmark, with a migration path to the standalone smolagents library. The work advocates code-driven actions over JSON, citing ~30% fewer steps and easier tool reuse via a Python-based interpreter, which can translate to lower LLM call costs for complex multi-tool workflows. By upgrading to smolagents, teams gain API-parity and a clearer upgrade path from transformers.agents, enabling faster iteration on agentic systems. This demonstrates a concrete capability uplift for automation pipelines that rely on external tools and multi-step reasoning.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- info