AI Potluck
Product / UX / Orchestration & agents

smolagents

Hugging Face

Minimal code-first agent framework from Hugging Face (~1,000 lines of core code) whose headline CodeAgent writes actions as executable Python rather than JSON tool calls, which research shows uses ~30% fewer steps than traditional tool-calling agents on hard benchmarks. Differentiates through model-agnostic design (Hugging Face Inference, OpenAI, Anthropic, local transformers/llama.cpp) and first-class MCP support, making it the path of least resistance for HF users building agents. 27.5K GitHub stars, 203 contributors, Apache-2.0; featured in Hugging Face's official agents course and referenced in Anthropic and Cohere agent-building guides.

smolagents by Hugging Face, minimal (~1k-LOC core) code-writing agent framework ('agents that think in code'). Apache-2.0. Verified live June 2026 on GitHub.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public(github.com/huggingface/smolagents)
core-gated
ungated

Fully OSI-licensed (Apache-2.0), entire library public, no proprietary/managed core, top openness, above open-core framework anchors (CrewAI/LangChain).

Adoption

3 high confidence
3.0

694,642 PyPI downloads in the last month (pypistats, confirmed). Monthly download volume in the 100K-1M band, meaningful but below heavyweight framework anchors (LangChain, CrewAI).

Capability

3 medium confidence
3.0

Deliberately lightweight: strong on model-agnosticism, tool-use, and sandboxing but thinner orchestration/eval surface than heavier frameworks; no benchmark, scored 3 on feature matrix.

Unchanged since 2026-07-30 (last edited, not re-checked)