Summary: Agentic Ai: Foundations, Architectures, And Applications

Read the summary and the most important questions on Agentic AI: Foundations, Architectures, and Applications

• 1 Define and classify AI agents according to established taxonomies, distinguishing them from traditional AI systems and chatbots.

  This is a preview. There are 91 more flashcards available for chapter 1
  Show more cards here

• Define an AI agent and explain what each of these elements means: perceives an environment, takes actions through actuators, goals, and autonomy.

  An AI agent is a system that perceives an environment, reasons about what it perceives, and takes actions (through actuators) to achieve goals.
  

  • Perceives an environment
    
    • Perception means receiving information from the environment.
  • Takes actions through actuators
    
    • An action is something the system does that may change the environment.
  • Goals
    
    • The agent’s behavior is directed toward goals (it acts in order to achieve them).
  • Autonomy
    
    • Autonomy means the system decides what to do without a human specifying every individual step.

• In the formal description of an agent, what does the function f: Percept∗ → Action represent, and what exactly does Percept∗ mean?

  • f: Percept^∗ → Action represents the agent function: it maps what the agent has perceived to an action.
  • Percept^∗ means the complete sequence of percepts the agent has received, not only the latest observation.

  
  So, the agent function chooses an action based on the agent’s percept history, not just the current input.

• What is the difference between the agent function and the agent program?

  • Agent function
    
    • The abstract mapping from Percept^∗ (percept history) to an Action.
  • Agent program
    
    • The concrete implementation of that function on hardware or software.

  
  In other words, the function describes what mapping the agent should implement; the program is how it is actually implemented.

• Using the criteria in the summary (ongoing perception-action process, autonomy, goal-directed behavior), explain why a pocket calculator is not normally considered an AI agent.

  A pocket calculator is not normally considered an AI agent because it lacks the key agentic properties:
  

  • No ongoing perception-action process: it primarily receives an input and returns an output, rather than continuously perceiving and acting in an environment.
  • Low autonomy: it does not independently decide what to do next; the user specifies each step/input.
  • Not goal-directed in the agent sense: it applies a calculation rather than pursuing goals via actions that may change an environment.

• Contrast a keyword-based spam filter with an agentic coding assistant in terms of perception, autonomy, and action in an environment.

  • Keyword-based spam filter
    
    • Perception: perceives incoming email.
    • Action: classifies email (e.g., spam/not spam).
    • Autonomy: usually follows fixed rules with little autonomy, so it is less agentic.
  • Agentic coding assistant
    
    • Perception: reads files and also reacts to outcomes (e.g., test results).
    • Actions: decides what to change, edits code, runs tests, and responds to results—actions that change its environment.
    • Autonomy: it decides what to do next across multiple steps, making it clearly agentic.

• An AI thermostat perceives only the current temperature and uses fixed condition→action rules to switch heating/cooling on or off. In the standard agent taxonomy, what type of agent is it, and what feature of its design determines that classification?

  It is a simple reflex agent.
  

  • Determining feature:
    
    • It uses only the current percept (current temperature).
    • It follows fixed condition–action rules (if temperature is below/above a threshold, turn heat/cool on/off).
    • It does not rely on memory or an internal representation of past percepts (no internal state is needed for the decision).
    
    

• A smart lighting system “remembers” that motion was detected ten minutes ago and can infer someone is still in the room even when no motion is currently detected. What agent type does this exemplify, and what role does internal state play in that classification?

  It exemplifies a model-based agent.
  

  • Role of internal state:
    
    • The system maintains an internal state that reflects more than the current percept (e.g., motion detected 10 minutes ago).
    • That internal state supports inference about the current situation even when the current sensor reading provides incomplete evidence (e.g., no motion now, but likely someone remains).
    
    

• A navigation system knows the current location, represents a destination as a goal, searches for a route, and changes the route when a road is blocked. Classify it in the agent taxonomy and justify the classification using its capabilities.

  It is a goal-based agent.
  

  • Justification:
    
    • It represents a destination as a goal.
    • It performs search to find a route that achieves that goal.
    • It can re-plan by changing the route when conditions change (e.g., a road is blocked), while still pursuing the same goal.
    
    

• A travel planner considers several routes to the same destination and compares them by time, price, comfort, or safety. What agent type is it, and what distinguishes it from a purely goal-based agent in this scenario?

  It is a utility-based agent.
  

  • What it does:
    
    • It evaluates how desirable different outcomes are using criteria such as time, price, comfort, and safety.
    
    
  • Distinction from purely goal-based (here):
    
    • A purely goal-based agent focuses on achieving the destination.
    • A utility-based agent additionally compares multiple goal-achieving options to choose the route with higher overall desirability.
    
    

• A coding assistant can inspect a repository, identify a programming goal, edit files, run tests, observe errors, and adjust its plan. Using the taxonomy language from the summary, explain why it can fit multiple agent categories at once, naming the relevant category-features.

  It can fit several taxonomy categories at once because it combines multiple category-defining features:
  

  • Internal state
    
    • It can maintain information gleaned from inspecting a repository and from test outcomes.
    
    
  • Goal-directed planning
    
    • It can identify a programming goal and plan edits/actions to achieve it.
    
    
  • Action and observation loop
    
    • It can edit files (act), run tests (act), observe errors (perceive), then revise what it does next.
    
    
  • Adaptation
    
    • It can adjust its plan in response to observed errors.