smolagents
A barebones library for agents that think in code -- the agent writes and executes actual code to take actions, rather than emitting structured JSON tool calls.
- Actively maintained
- Well documented
- Permissive license
Overview
smolagents takes a code-first approach: instead of an agent emitting a JSON blob that a framework then interprets as a tool call, the agent writes real code (typically Python) to take actions directly, which Hugging Face's research suggests is often more capable and efficient than JSON-based tool calling for complex tasks. The library itself is intentionally minimal.
smolagents is a fit for teams who want to experiment with or adopt the code-execution paradigm for agents rather than the more common JSON-tool-call pattern used by most agent frameworks (several already in this catalog) -- a genuinely different approach, not just another framework with the same underlying model.
- Categories
- AI Agents & Agent Tools
- Languages
- Python
- License
- Apache-2.0
Related Projects
LangGraph
MITBuild resilient, stateful agents as graphs -- explicit nodes and edges for control flow, rather than an implicit agent loop, for reliable long-running or multi-step workflows.
Actively maintained
- AI Agents & Agent Tools
Composio
MITPowers 1000+ toolkits, tool search, context management and authentication so AI agents can actually take action, not just plan it.
Actively maintained
- AI Agents & Agent Tools
Letta
Apache-2.0A platform for building stateful agents with advanced memory that learns and self-improves over time, rather than resetting between conversations.
Actively maintained
- AI Agents & Agent Tools
MCP Servers
MITOfficial reference implementations of the Model Context Protocol -- the open standard now adopted across OpenAI, Google, Microsoft and Anthropic for connecting agents to tools and data.
Actively maintained
- AI Agents & Agent Tools
Spotted an error? Suggest an edit on GitHub.