Introduction
An assistant becomes more consequential when it can choose steps and use tools. Before calling it an agent, ask what it can observe, decide, and change. What are its limits? How would we know it succeeded? Answer those questions using evidence about the specific system.
Building on application architecture in Module 20 and decision authority in Module 17, we will examine workflows, planning loops, tools, state, orchestration, MCP, human review, evaluation, and recovery. Your project specifies a research agent that investigates a question, analyzes and compares information, and prepares a briefing. You will also deliberately break it and explain what happened in a postmortem.
Start with control loops, then interfaces and permissions, then operating examples. Use fictional data for the practice. Keep tools read-only apart from creating local drafts, so you can investigate failures without changing real external records.