🛠️ Lab 1 — Build and Benchmark Your First Code-Writing Agent (50 pts)

Goal: Build a working code-writing agent using Hugging Face's free smolagents library, then extend it to run against three different free/local model providers so you can directly observe how the provider landscape affects latency, cost, and behavior — using only free or local tools. Steps:

1. Open a free Google Colab notebook (or a local Jupyter environment if you prefer). Install the required library: pip install smolagents duckduckgo-search.
2. Create a free Hugging Face account if you don't already have one, and generate an access token (Settings → Access Tokens) for the Serverless Inference API.
3. Follow the structure of the Hugging Face Agents Course tutorial ("Build Your First Agent with smolagents," linked in Resources) to instantiate a CodeAgent using the Qwen2.5-Coder-32B model via the free Serverless Inference API, and equip it with the DuckDuckGoSearchTool.
4. Give your agent a concrete, multi-step task that requires both reasoning and a web lookup — for example: "Look up the current population of the three largest cities in California and write a Python function that returns them sorted in descending order." Run it and read the full trace of tool calls and reasoning steps it produces.
5. In a markdown cell, annotate the trace: identify each point where the agent plans, acts, observes, and adapts. Label at least three distinct loop iterations.
6. Now extend the exercise. Point the same agent logic at two additional free providers so you can compare across the provider landscape:
• Groq (free tier, no credit card required, 30 requests/minute, 14,400 requests/day) — sign up for a free API key and swap in a Groq-hosted open model.
• Google Gemini free tier (Flash-Lite model, 15 requests/minute, 1,000 requests/day) — sign up for a free Google AI Studio API key.
• (Optional stretch): install Ollama locally and run a Qwen 3 model entirely on your own machine at zero API cost, for a true local baseline.
7. Run the identical task against each provider. For each run, record: wall-clock latency (seconds from prompt to final answer), estimated cost (should be \$0 for all three), and any differences in output quality, tool-call count, or failure modes.
8. Compile your results into a short comparison table and a 200-300 word written reflection on what the exercise revealed about the tradeoffs between providers.

Deliverables: Submit your annotated Colab notebook (as a downloaded .ipynb or PDF export) plus your comparison table and written reflection, either pasted into the text entry box or attached as a separate document.