Source: smolagents API reference: Agents (v1.26.0)

Constructor parameters and defaults for MultiStepAgent, CodeAgent and ToolCallingAgent, run() options, AgentMemory, prompt templates and save/push_to_hub.

Contents

Summary

Reference for the agent classes. It labels smolagents an experimental API subject to change.

Key claims

  • MultiStepAgent: max_steps default 20, planning_interval, instructions, managed_agents, step_callbacks, final_answer_checks, name/description (for managed agents), provide_run_summary, return_full_result (default False), add_base_tools (default False).
  • CodeAgent: additional_authorized_imports, executor_type in local (default), blaxel, e2b, modal, docker; executor_kwargs; max_print_outputs_length; stream_outputs; use_structured_outputs_internally (structured generation per step, default False); code_block_tags.
  • ToolCallingAgent: uses the model’s native tool calling; max_tool_threads for parallel tool calls.
  • run(task, stream, reset, images, additional_args, max_steps, return_full_result); additional_args injects Python objects (e.g. dataframes) into the run.
  • AgentMemory holds the system prompt and TaskStep / ActionStep / PlanningStep list; return_full_code() concatenates all code actions; replay() pretty-prints.
  • PromptTemplates customise system, planning, managed-agent and final-answer prompts.
  • save() writes tools, managed agents, agent.json, prompt.yaml, app.py and requirements; from_hub requires trust_remote_code=True because it downloads and runs code.

Feeds

Suggest changes →