smolagents

You will like smolagents because there is almost nothing to it: the agent logic is about a thousand lines, so when it misbehaves you read the source instead of the docs, and the fix is usually yours to make in an hour. That is the daily trait that decides it. It goes as far as a single agent loop goes and no further; there is no persistence, no scheduler and no opinion about your application.

Pick it for research, scripts, teaching and anything you want to understand completely. Pick LangGraph when the workflow has to survive a restart, and the OpenAI Agents SDK if you want handoffs without writing them.

The risk is the executor. The design is a model writing Python that is then run, and isolation is something you add: a CodeAgent without a configured sandbox runs generated code where you are standing, with your credentials in the environment. Code actions are more expressive than JSON tool calls, and more expressive means a wider blast radius when the model is wrong.

The sandbox argument is mandatory in anything that touches real data. What it does right: four sandbox backends are documented, Modal, Blaxel, E2B and Docker, so the safe path exists and is one constructor argument away.

1. CodeAgent emits a code block per step, so a single action may nest calls, loop and branch, where JSON tool calling needs one round trip per call; the documented justification is composability, and it is a real efficiency argument, since fewer round trips means fewer tokens per useful action. 2. ToolCallingAgent offers the conventional JSON path for models that prefer it. 3. Inputs may be text, image, video or audio.

No benchmark is cited, and no verification beyond the executor's own errors is described. The observation: the framework's thesis fits in a sentence, which is more than most frameworks can say.

No pricing page, no plan, and that is the strategy. InferenceClientModel with no arguments routes to Hugging Face's Inference providers, so every tutorial that runs is a metered call to the marketplace, and roughly 29,000 stars is a lot of tutorials. Moat: the Hub's gravity, which the library exists to increase rather than to monetize directly.

Likely outcome: it stays alive as long as the Hub does, and the Hub is the parent's whole business. No acquirer; the parent is the exit. Position: long the parent, neutral on the library, and set your model class explicitly.

The demo is five lines. As a dependency: maintained by Hugging Face, $0 for sixty engineers, no SLA, and issues go to GitHub. The concern is the Hub integration: agents and tools can be shared as Gradio Spaces, so an engineer can publish internal tooling, prompts included, with one call and a token that was meant for downloads.

CI fit is a pip install. Onboarding is a morning. Approved with conditions: Hub publishing blocked at the org level, tokens scoped to read, and the executor reviewed before anything runs on a shared runner.

This one I own. Apache-2.0, pip install 'smolagents[toolkit]', and TransformersModel runs weights on my own GPU with no network, or Ollama if I prefer the daemon. Tools come from ToolCollection.from_mcp for any MCP server or Tool.from_langchain for anything the other ecosystem already wrote, so the tool shelf is everyone's shelf.

A fork is trivial; the whole thing is small enough to vendor into a project and forget it is a dependency. The only reason to fork would be taste, and the code is clean enough that taste would be the only complaint.