🤖 Agent Basics
What is the key difference between a regular LLM and an Agent?
An LLM gives one response and stops; an Agent loops through Think→Act→Observe→Think until the task is done.
What does ReACT stand for?
Reasoning + Acting (Thought → Action → Observation cycle).
Name the 4 Multi-Agent architecture types we discussed.
Sequential (pipeline), Hierarchical (boss/worker), Parallel (with aggregator/voting), and Debate/Critic.
What is a "tool" in the context of an Agent, conceptually?
A defined function or capability that extends what the Agent can DO beyond generating text
Name one real-world product or system mentioned in our course that already uses multiple specialized "agents" working together behind the scenes.
Google Search, self-driving car, etc.
Name the 4 core components of an Agent's architecture (any order).
Core LLM (brain), Tools/Actions, Memory, and Planning/Reasoning (Observation/Feedback also acceptable as a 5th).
Put these 4 steps in the correct order: Observe, Act, Reflect/Think, Reason.
Reason (Think) → Act → Observe → Reflect/Think (then loop back to Reason if needed).
Which Multi-Agent architecture would be BEST for a major investment decision where you want multiple perspectives debated before reaching a conclusion? Why?
Debate/Critic architecture — because it's designed for high-judgment decisions where considering multiple viewpoints leads to a more well-rounded, carefully considered conclusion.
Name 3 characteristics of a well-designed tool.
A clear/specific name, a small number of well-typed parameters, clear documentation, structured (e.g., JSON) return values, and proper error handling.
What is MCP, and conceptually, what problem is it trying to solve for Agents?
MCP (Model Context Protocol) is a standardized way for Agents to connect to external tools, data sources, and services. It solves the problem of having to build a custom, one-off integration for every single tool/service an Agent might need — providing a common "plug" format instead.
Why might using an Agent for "translate this one sentence" be a waste of resources?
It's a simple, single-step task — a regular LLM can do it in one shot. An Agent's looping/tool-calling overhead adds unnecessary cost and time for something that doesn't need multiple steps.
What is "Context Window pressure" and why does it get worse as a Single Agent takes more steps?
Context Window pressure is when the Agent's memory/input space fills up with all the accumulated information from previous steps. As more steps happen, more info piles up, making it harder for the Agent to remember earlier constraints or details
As an Agent takes more and more steps in a ReACT loop, what TWO problems tend to get worse, and why?
(1) Context window pressure
(2) Error compounding
Why is it generally better for a tool to return structured data (like JSON) instead of a long block of free-form text?
Consistent and predictable
Give one general reason why giving an AI/Agent unrestricted permissions/resources without any limits can lead to unintended, costly outcomes — without referencing any specific story we discussed.
An Agent optimizes purely for completing its given goal as efficiently as possible; without limits, it may interpret "efficient" as using as many resources as needed/available, even if that scale is wildly disproportionate to what a human would consider reasonable for the task.
What does it mean for an Agent to have "memory," and why is this different from a regular LLM's behavior within a single conversation?
Memory means the Agent can store and recall information across multiple steps — like results from earlier tool calls or facts it has gathered — and use that stored information to inform later decisions, rather than just responding to the immediate prompt in isolation.
After the "Observation" phase, what does the Agent do next, and what determines whether it loops again or stops?
It reflects/thinks about whether it has enough information to complete the goal. If yes, it produces the final output; if no, it loops back to "Thought" to plan the next action.
Compare Hierarchical and Sequential architectures: which one is more ADAPTIVE to unexpected situations, and why?
Hierarchical is more adaptive
What are the 5 parts of a well-structured Agent Skill?
Name+Trigger, Goal, Inputs, Steps, Output Format.
Why do AI researchers generally recommend keeping a human "in the loop" for major or irreversible Agent decisions (like deleting data or spending large sums of money), even as Agents get more capable?
Because Agents can be very good at executing a defined goal efficiently but are not good at judging whether an action is appropriate, proportionate, or safe in a broader human context. That's where human should be responsible for!
Explain the relationship between an Agent's "Planning" component and its "Tools" component — how do they work together during task execution?
The Planning component decides WHAT needs to happen next and in what order (the strategy), while the Tools component is HOW the Agent actually executes a specific step of that plan (the capability). Planning determines which tool to call and when; Tools provide the actual mechanism to act on the world and get results back.
Why is the ReACT loop considered "self-correcting" compared to a single LLM response?
Because at each Observation step, the Agent can detect if something unexpected happened and adjust its next Thought/Action
In a Sequential architecture, what is the main risk if one agent early in the pipeline produces a low-quality result?
Every downstream agent inherits and builds on that flawed result
Design challenge: explain why a tool named do_something(x) is a BAD tool design?
1) Vague name — doesn't tell the Agent what it does; (2) Unclear parameter (x has no descriptive name or type); (3) No documentation/docstring explaining usage; (4) Likely no structured return value; (5) The Agent won't know when or how to use it correctly, leading to errors or misuse.
Imagine you're designing an Agent system for a school to help automatically grade essays and email feedback to students. Using concepts from AT LEAST 3 different categories in this game (Fundamentals, ReACT, Multi-Agent, Tools/Skills, or Practice), explain how you would design this system to be both effective AND reliable.
Anything that makes sense!