Building AI Agents with smolagents

Understand large language models and their limitations, then see how AI agents extend LLMs with tools like web search to interact with environments and achieve goals.

An AI agent fuses a language model brain with a tool-enabled body to act toward a goal, where the brain reasons and routes and the body uses tools.

This lecture contrasts AI agent workflows and real agents, explaining how developers shape workflows from prompt chaining to control flow, while real agents act autonomously with environment feedback.

Use workflows for critical situations with defined structures. Use agents when outcomes are unpredictable and require creativity or accommodate diverse user requirements.

Explore multi-agent systems built from specialized agents with dedicated prompts and tool sets. The manager agent delegates tasks to web and coding agents and compares parallel, sequential, and competitive modes.

Compare popular frameworks for small agents, such as CRI, Autogen, Lang, Chain, Landgraf, and LlamaIndex, and apply Anthropic's guidance on when to use agents and prompt chaining.

smolagents offers a lightweight library to build agents with a language model and tools, featuring a code-first Python approach, built-in tooling, and gradio-based visualization.

Create your first agent with smolagents by initializing a uv-powered Python project, adding dependencies, and running main.py. Authenticate to hugging face with an access token stored as hf_token.

Observe how a multi-step agent builds a plan, uses web research and Python code to execute action steps, and logs results in the terminal for transparency.

Compare code-based actions and json tool calls in multi-step agents, and learn why code agent actions offer better composability, handling of complex objects, and generality.

Build a weather agent with a custom tool to fetch Tokyo weather data from sample data, import matplotlib for bar chart visualization, and execute Python code safely in a sandbox.

Explore how the code agent executes Python code locally and in secure cloud sandboxes for production deployment, with a default safe import list and optional additional imports for trusted libraries.

Learn to add a gradio-based user interface to a smolagents weather agent, launch the UI, access the local URL, and interact through prompts with console log details.

Publish your smolagents to the Hugging Face hub spaces, test them with a Gradio UI, and push or grab agents and tools from the hub.

Build a multi-agent system that orchestrates web search, content analysis, and final market reports via a manager delegating to a web search agent and an analyst.

Monitor ai agents in production with opentelemetry and Langfuse, exploring traces and spans for visibility. Learn local setup with docker and instrument the agent to diagnose latency and model usage.

Learn to set up and use MCP servers locally with small agents, discover servers via Smithery, configure environments and API keys, and inspect tools with the inspector.

Explore ai agents with small agents, from large language model limits to brain and body architecture, workflows, multi-agent systems, weather and code agents with tool calling.