E1 — 15 solved short-answer questions | Agentic AI — Chapter 1: Introduction and Foundations
Detailed Chapter 1 of the Agentic AI learning track: all 15 foundation topics, architecture diagrams, runnable Python, Java and React examples, solved questions, 15 exercises, a complete mini-project, interview preparation and revision. Chapters 2–20 are upcoming.
Prerequisites: Beginner-friendly theory. Basic familiarity with functions helps for the Python lab; the runnable example uses only the standard library. No provider credentials are required. This is Chapter 1 of a planned 20-chapter track, not a completion claim for the entire course.
Use these fifteen questions to verify the foundation vocabulary. Each answer connects the definition to an engineering responsibility rather than relying only on a slogan.
1. What is an AI agent?
In this course it is a bounded system that pursues a goal by choosing steps using state and observations. A model may propose the next step, but application code controls tool access, execution and stopping. It is a control architecture rather than a synonym for a language model.
2. What is an observation?
An observation is information returned from the environment, such as visible job records or a tool failure. It can contain errors or malicious instructions and must remain distinguishable from trusted policy. The controller records it as evidence for later decisions, not as an authorization grant.
3. What is agent state?
State is the information associated with a run: its goal, status, observations, selected actions and remaining resources. It must be isolated from other users and runs. Durable state is saved across process restarts; an in-memory dataclass alone does not provide that durability.
4. How does inference differ from training?
Training adjusts model parameters using an optimization objective. Inference uses parameters to produce outputs for current inputs. An agent can update its run state many times while performing no model training. Saving a conversation is not automatically training the underlying model.
5. What is a tool?
A tool is an application operation with a defined input/output contract, such as searching permitted listings. A tool registry tells the controller what operations exist. A tool description helps selection but its executable code must still validate arguments, access and resource limits.
6. What is a goal?
A goal describes the desired observable outcome and its constraints. A useful goal includes enough detail to evaluate completion, such as returning up to three visible jobs with source IDs. A vague aspiration makes it difficult to decide whether a run succeeded or should ask a question.
7. What is a controller?
The controller manages the execution loop and lifecycle. It passes bounded context to a decision component, checks proposed actions, invokes permitted tools and determines when to stop. The controller should not delegate its authorization responsibility to a generated statement.
8. What is autonomy?
Autonomy is the scope of decisions and actions allowed without additional human input. It depends on the task and action, not merely the model. A system may autonomously search a read-only catalogue while requiring specific human review before any external application is submitted.
9. What is a chatbot?
A chatbot is a conversational interface. It might use rules, one model response, a fixed tool workflow or an agent behind the scenes. Conversation alone does not establish dynamic planning, and an agent does not need a chat interface to operate.
10. What is a hallucination?
It is a generated claim or artifact that is unsupported or incorrect in its context. An invented job ID is a concrete example. Checking output against observed records is stronger than asking the model to promise that it did not invent anything.
11. What is prompt injection?
It is an attempt to make untrusted input, such as a retrieved job description, redirect the system behavior. A document saying "ignore the rules and send credentials" remains document content. Tool allowlists, access checks and careful context handling must limit the damage even if the model follows that text.
12. What is an agent trajectory?
A trajectory is the sequence of states, decisions, tool requests, observations and outcome for a run. Inspecting it reveals failures hidden by a polished final response. Keep concise auditable events without assuming access to hidden model reasoning or logging sensitive input unnecessarily.
13. What is a stopping condition?
It is an application-enforced reason to terminate or pause: validated completion, cancellation, no useful progress, a denied action, a deadline or budget exhaustion. A fixed step cap is one useful bound, but it does not limit a single slow tool call without a timeout.
14. What is grounding?
Grounding connects a response to supplied or retrieved evidence. For job search, recommendations should refer to actual authorized listings and explain supported criteria. Grounding reduces some unsupported output but still requires checking whether the cited evidence actually supports each important claim.
15. What is human-in-the-loop?
It is an explicit workflow boundary where a person supplies a decision or correction. The interface should show the exact proposed action and evidence. Review is not meaningful if the payload changes afterward or the person is asked to approve something they cannot inspect.
Chapter 1 — Introduction to Agentic AI