smolagents

Each agent runs a ReAct-style loop: the model reads the task plus prior steps, proposes an action, the framework executes it and appends the observation, and the loop stops when the agent calls final_answer or hits max_steps. In a CodeAgent the action is a Python snippet run by an interpreter or remote executor; in a ToolCallingAgent it is a JSON tool call. An optional planning_interval inserts periodic planning steps.

Multi-agent work is hierarchical. A sub-agent gets a name and description and is passed to a manager through managed_agents; the manager then calls it much like a tool and receives its result. There is no shared blackboard or graph definition.

State lives in agent.memory, an ordered list of typed steps (system prompt, task, planning, action). Calling run(..., reset=False) keeps that memory for a follow-up run. No built-in long-term or cross-session store is documented.

The docs show human review through step_callbacks: a callback registered for PlanningStep pauses the run after a plan is written so a person can approve, edit or cancel it, and run(task, reset=False) resumes with memory intact. A running agent can be stopped with agent.interrupt() (the guided tour suggests wiring it to a button in GradioUI), and memory steps can be edited before the next step. No built-in per-tool approval gate is documented.

pip install "smolagents[toolkit]"

from smolagents import CodeAgent, ToolCallingAgent, InferenceClientModel, WebSearchTool

model = InferenceClientModel()  # uses HF_TOKEN from the environment

searcher = ToolCallingAgent(
    tools=[WebSearchTool()],
    model=model,
    name="searcher",
    description="Looks things up on the web and reports short findings.",
    max_steps=6,
)

lead = CodeAgent(tools=[], model=model, managed_agents=[searcher])
print(lead.run("In which year was Python 3.0 released, and how many years before 2026 was that?"))

• Requires Python 3.10 or newer.
• InferenceClientModel needs an HF_TOKEN for Hugging Face Inference Providers; other providers need their own keys.
• Managed agents must have both name and description, or the manager cannot call them.
• CodeAgent only allows whitelisted imports; add more with additional_authorized_imports.
• The default LocalPythonExecutor is not a security sandbox; use Blaxel, E2B, Modal or Docker for untrusted code.
• MCP tools need the extra: pip install "smolagents[mcp]".
• With Ollama through LiteLLMModel, the guided tour warns that Ollama's default 2048-token context fails and sets num_ctx=8192.

Yes, as a client. With the mcp extra installed, MCPClient loads tools from stdio or Streamable HTTP MCP servers and hands them to an agent. A2A and AG-UI support were not found in the official docs.

The library is Apache-2.0 licensed with no paid tier of its own. Model calls are billed by whichever provider you configure; Hugging Face Inference Providers need an HF_TOKEN.

Python. The installation guide requires Python 3.10 or newer, and no SDK for other languages is documented.

Through a hierarchy: sub-agents with a name and description are passed to a manager agent via managed_agents, and the manager calls them as it would call a tool.

The project warns that LocalPythonExecutor is not a security boundary. For untrusted code the docs recommend Blaxel, E2B, Modal or Docker executors.