Module Check: Architecture Matters
This module opened with a claim: an AI agent is an architecture around a model, not the model alone. Over nine lessons you have examined every component of that architecture in depth — the LLM reasoning core's capabilities and limits, the planner's goal-decomposition strategies, the tool layer's three-part structure and design principles, the memory system's four types and management strategies, and the orchestration loop's canonical cycle and failure modes. You have seen these components realized in real frameworks and traced them through a concrete worked example. This final lesson consolidates that knowledge and asks you to synthesize it.
Flashcards — click each card to reveal the answer
Module Quiz: All Five Components
An agent is given the goal: 'Monitor our AWS cost dashboard, identify the three services with the highest cost increase this month, and draft a Slack message summarizing the findings.' Which component converts this single goal sentence into a step-by-step execution plan?
A production agent has been running for six months. Users report that it gives factually incorrect answers about events from the last three months. Which LLM core limitation is most directly responsible, and what is the correct architectural fix?
Two tools in an agent's tool layer are named get_data (description: 'Gets data') and fetch_records (description: 'Fetches records'). An engineer notices the agent frequently selects the wrong tool. Which tool design principle is most clearly violated?
An orchestration loop has no error handling for tool failures. A web search tool returns a 503 error. What is the most likely consequence?
A student argues: 'If I use a more powerful LLM core, I don't need a well-designed memory system or tool layer — the model will figure it out.' What is the most accurate rebuttal?
In the ReAct pattern, what is the purpose of the Thought: prefix before each Action?
Capstone: Agent Design Defense
- This capstone synthesizes the full module. You will design an agent, justify every architectural decision, and present it to be challenged.
- SCENARIO
- You are an AI engineer at a nonprofit that provides free tutoring to under-resourced high school students. Your team wants to build an AI tutoring agent that can: answer subject-matter questions in math, science, and English; quiz students adaptively (harder questions when they get answers right, easier when they get them wrong); track each student's mastery of specific topics across sessions; and flag students who are falling behind to a human counselor.
- TASK
- Design the full agent architecture for this tutoring system.
- Step 1 — FIVE COMPONENTS SPEC
- For each component, write 2-4 sentences:
- - LLM core: model choice and justification (why this model for this use-case?)
- - Planner: what planning strategy (one-shot, hierarchical, reactive)? Why?
- - Tool layer: list all tools with type and description. Which are retrieval, action, computation?
- - Memory system: what memory types are used? What is stored and when? How is the context window managed across a long tutoring session?
- - Orchestration loop: what stopping conditions? What human-in-the-loop checkpoints? How are errors handled?
- Step 2 — FAILURE MODE ANALYSIS
- Identify three failure modes specific to the tutoring use-case (not just generic agent failures). For each, name the component responsible and propose a specific mitigation.
- Step 3 — ETHICAL CONSIDERATIONS
- This agent interacts with minors. Identify two ethical considerations specific to this use-case (data privacy, bias in adaptive difficulty, dependency on AI, etc.) and describe how your architecture addresses each.
- Step 4 — DEFENSE
- Share your design with a small group. Each group member asks one 'what if' challenge question (e.g., 'What if a student asks the agent to do their homework for them instead of tutoring them?'). Revise your design to address the challenges raised.
- Deliverable: A written 5-component spec, 3 failure mode analyses, 2 ethical considerations, and documented revisions from the peer defense.