Project 3: Thought Cascade

Building and Evaluating Agentic AI Systems with Small Language Models

Overview

In this assignment, you will explore the transition from traditional “single-shot” prompting to Agentic AI. You will build a system that uses the ReAct (Reasoning + Acting) pattern to solve complex debugging problems by iterating, testing, and self-correcting.

Learning Objectives

By completing this project, you will:

  • Understand the theoretical foundations of agentic AI and why iteration beats size
  • Build complete agentic systems using the ReAct pattern
  • Measure system performance at the component level, not just end-to-end
  • Design tools that enable effective agent-environment interaction
  • Diagnose failure modes and implement recovery strategies
  • Apply agentic patterns to diverse problem domains

Materials

TipQuick Access

Project Notebook (GitHub) — Complete implementation guide and assignment
Setup Guide — Local setup instructions

ImportantSubmission Requirements
  1. Complete all coding implementations marked with TODO in the notebook.
  2. Complete all analysis sections explicitly labeled “Required Analysis”.
  3. Run all cells so that key tables and plots are visible in your saved notebook.
  4. Follow course submission instructions for exporting your answers.

What You’ll Build

Core Agentic Components

You’ll implement the building blocks of an autonomous agent: - Code Execution Environment: A safe sandbox for the agent to run and test code. - ReAct Loop: The main control loop that interleaves Reasoning, Acting, and Observing. - Tool Use: Mechanisms for the model to invoke external functions.

Evaluation Frameworks

You’ll build rigorous ways to measure performance: - Pass@k Metrics: Measuring reliability across multiple attempts. - Strict Testing: Creating adversarial test cases to catch subtle bugs. - Trajectory Analysis: Evaluating the quality of the agent’s reasoning process.

Technical Stack

  • Models: Gemma-2-2b-it (Google) or SmolLM-360M (HuggingFaceTB)
  • Frameworks: PyTorch, Hugging Face Transformers
  • Dataset: QuixBugs (classic algorithmic bugs)
  • Compute: CPU (optimized for Apple Silicon/standard CPU) or GPU

Research Context

This project connects to the forefront of AI research:

  • ReAct (Yao et al., 2022) — Synergizing reasoning and acting in language models.
  • Agentic Workflows — Moving beyond simple QA to autonomous problem solving.
  • Small Language Models — Showing how agentic patterns allow smaller models to outperform larger ones on complex tasks.

Previous: ← Project 2: Neural Archaeology | Next: Capstone Project →