Hugging Face just shipped Transformers Agents 2.0, and the headline number is hard to ignore: a Llama-3-70B-Instruct agent built with the new framework now outperforms GPT-4-based agents on the GAIA leaderboard. That’s a meaningful shift, not just a benchmark flex. It suggests the agent scaffolding—how you structure tool access, iteration, and memory—can matter as much as raw model capability.
The release introduces two new agent types alongside the existing CodeAgent. ReactCodeAgent generates actions as Python blobs and can iterate on previous observations, while ReactJsonAgent does the same with JSON output. The core loop is familiar to anyone who’s worked with ReAct-style systems: Thought, Action, Observation, repeat until the task is solved. What Hugging Face is emphasizing isn’t novelty so much as clarity. The framework keeps abstractions minimal, exposes attributes like the final prompt and tools, and separates the agent logic from the underlying LLM so you can swap models freely.
There’s also a community angle. Agents can now be shared, which lowers the barrier for teams that don’t want to build every tool from scratch. The toolbox concept—pre-instantiated tools ready to go—addresses a real pain point: initialization time. If you’ve ever watched an agent spin up tools for 30 seconds before doing anything useful, you’ll appreciate the design choice.
One thing worth watching: the framework has already been spun out into a standalone library called smolagents, with a nearly identical API. That’s a signal Hugging Face expects this to grow beyond the transformers package. The v4.41.0 release lands mid-May, but you can install from the main branch now if you want early access. Given how quickly agent frameworks are evolving, the GAIA result won’t stay unusual for long—but it’s a useful data point for anyone still assuming GPT-4 is the default choice for agent workloads.
