This module dives into the Reflection design pattern and shows how it can significantly improve the quality and reliability of LLM-based workflows. You will contrast direct (zero-shot) generation with reflection-driven approaches where an LLM critiques and revises its own outputs or responds to external feedback such as error messages or evaluation signals. Through concrete examples—like revising emails, debugging code, and improving chart visualizations—you will see how to construct prompts that ask models to review drafts against explicit criteria and then produce better second versions. The module covers multimodal reflection, where an LLM critiques images or charts, and introduces rigorous evaluations to quantify the gains from reflection, such as improved SQL accuracy in database-query tasks. You will learn how to incorporate runtime signals and external tools into reflection loops, how to design reflection prompts and workflows that outperform simple prompt tuning, and how to recognize where reflection is most beneficial.

This module focuses on tool use and function calling, which allow LLMs to access external capabilities such as APIs, databases, and code execution. You will learn what tools are in the context of agentic workflows and how to expose functions as tools that models can request when needed. The module walks through older marker-based approaches and modern native tool-calling syntax, emphasizing patterns for detecting desired function calls, executing them, and feeding results back to the model. You will implement tools using libraries like AI Suite, explore an email assistant workflow that composes multi-step tool use, and see how code execution can be safely integrated so LLMs can write and run code to solve tasks. Finally, you will be introduced to the Model Context Protocol (MCP), a rapidly growing standard that standardizes access to external tools and data sources and greatly simplifies integrating agentic applications with diverse back-end services.

This module provides practical, step-by-step strategies for evaluating, debugging, and improving agentic AI workflows. Building on earlier coverage of evals, you will learn how to perform detailed error analysis by examining traces and failure cases in applications such as invoice processing and customer email response. The module teaches how to distinguish between end-to-end and component-level evaluations, how to isolate and tune specific components (for example, web search in a research agent), and how to use metrics such as precision, recall, and F1 where appropriate. You will explore techniques for optimizing both LLM and non-LLM components, including prompt design, model selection, hyperparameter tuning, and provider swaps, while balancing quality, latency, and cost. Through a structured build-versus-analyze workflow and a 2×2 framework for evaluation types, you will learn how to iterate effectively on agentic systems, prioritize high-impact fixes, and benchmark performance as you move toward reliable, production-ready deployments.

This module explores advanced design patterns for building more autonomous agentic systems that can plan, coordinate, and act with minimal hard-coded control. You will learn planning patterns in which LLMs generate executable plans—either as structured JSON or code—that orchestrate tools and data sources to accomplish complex tasks. Through examples such as retail agents and spreadsheet analytics, you will compare JSON-based plans with code-as-action approaches and understand their trade-offs in flexibility, security, and predictability. The module then turns to multi-agent workflows, where specialized agents with different roles and tools collaborate, and introduces communication patterns such as linear pipelines, hierarchical manager–worker setups, and all-to-all messaging. Finally, it revisits the performance, parallelism, and modularity benefits of agentic workflows, showing how autonomous agents can parallelize sub-tasks and leverage interchangeable components, before concluding with guidance on applying these patterns responsibly in real-world projects.