Qlaudia

Dark mode

Understanding Agentic AI: Concepts and Applications

Artificial Intelligence
Technology & Computing
Philosophy & Ethics

Agentic AI systems pursue goals autonomously by reasoning, taking actions through tools, and adapting plans based on observed results—unlike conversational AI that responds one exchange at a time. This course develops the vocabulary and mental models to understand how these systems work, their current capabilities and limitations, and how prominent frameworks and real products implement agentic workflows.

What you'll learn

The Agent Execution LoopThe repeating cycle of perceiving the situation, reasoning about it, planning, acting, and observing results that forms the core operational pattern of an agentic system, continuing until the goal is achieved or a stopping condition is reached.
Agentic AI Frameworks as a CategoryWhat an agentic AI framework or toolkit is in general terms, and why developers use such frameworks to structure the planning, tool use, and coordination logic of an agent rather than building it from scratch.
Multi-Agent OrchestrationHow complex agentic systems coordinate multiple specialized sub-agents, often through an orchestrator that delegates tasks and combines results, enabling parallel exploration and division of labor.
Tool Use and External ActionsHow an agentic system extends its capabilities beyond generating text by invoking external tools and resources, such as web search, code execution, file access, or APIs, to gather information or effect change in the world.
Defining an AI AgentWhat it means to call an AI system an 'agent': a system with a goal, the ability to make decisions, take actions in the world, and operate with some degree of autonomy, along with the recognition that agency is a spectrum rather than a strict binary.
Planning and Re-planningHow an agent drafts an approach to achieve its goal, either step-by-step or as an upfront full plan, and how it revises that plan when actions fail, produce unexpected results, or circumstances change.
Practical Limitations of Agentic AIThe recognized shortcomings of current agentic systems, including unreliable performance outside controlled demonstrations, non-deterministic behavior across runs, and difficulty sustaining performance in higher-risk real-world settings.
Current Capabilities of Agentic AIA general picture of what agentic AI systems can convincingly do today, including the kinds of multi-step, tool-using tasks where they perform well under controlled conditions.
Memory in Agentic SystemsThe role of persistent state and memory in allowing an agent to retain information across multiple steps or sessions, distinguishing it from the stateless, single-turn nature of ordinary conversational prompting.
Real-World Application: ChatGPT AgentAn examination of OpenAI's computer-using agent that navigates websites, fills forms, and completes tasks like bookings, including where it still requires human oversight, and how its operation reflects the agentic workflow.
Reasoning and Acting (ReAct) PatternThe influential pattern in which a model alternates between an explicit thought about what to do next, an action such as a tool call, and an observation of the result, using each observation to inform the next reasoning step.
Trends Shaping Agentic AI's TrajectoryThe near-term directions in which agentic AI is developing, including convergence across major providers on similar loop-based architectures and continued innovation in context management, coordination, and auditability.
Observing and Incorporating FeedbackHow an agent evaluates the outcome of an action it has taken, determines whether it succeeded or failed, and uses that observation to decide its next step, forming the feedback portion of the agent loop.
Real-World Application: ManusAn examination of the Manus general-purpose agent product, which plans and executes multi-step workflows such as research and report generation by running code in a sandbox, and how it exposes the plan-act-observe cycle to users.
Cost and Resource Trade-offs of Agentic AIWhy agentic workflows consume substantially more computational resources and tokens than single-turn chat, and how this shapes practical decisions like caching, budgeting, and monitoring in deployed systems.
Conversational LLMs vs Agentic SystemsThe fundamental distinction between an LLM that generates one response to a prompt and stops, versus an agentic system that persists across many steps, takes actions, and pursues a goal without constant human re-prompting.
Stopping Conditions and Runaway LoopsWhy agents without limits can loop indefinitely on a failing action, and the practical safeguards such as maximum iteration counts, no-progress detection, and cost or token budgets used to keep an agent's execution bounded.
Goal Interpretation and Task DecompositionHow an agentic system takes a high-level objective given by a user and breaks it down into a sequence of smaller, actionable sub-tasks it can work through.
Challenges in Evaluating Agentic AIWhy measuring agentic AI performance is difficult, including the lack of a consensus definition of agency, the inadequacy of static benchmarks for capturing real-world variability, and the risk of metrics being gamed rather than reflecting genuine capability.
LangGraph Framework OverviewA conceptual look at LangGraph, which represents agent workflows as stateful, directed graphs of steps and execution paths, and the kinds of complex, cyclical tasks it is suited for.
AutoGen and CrewAI Frameworks OverviewA conceptual look at Microsoft's AutoGen, which structures multi-agent collaboration as role-playing conversations, and CrewAI, which assigns agents fixed roles in a sequential pipeline resembling a human team.

About this course

What This Course Is About

You already know how to have a conversation with an AI system: you type a prompt, it responds, and the exchange either ends there or continues as a back-and-forth dialogue you steer at every turn. This course is about a different, and newer, way of using the same underlying technology — agentic AI. Instead of answering one question at a time, an agentic AI system is given a goal and left to figure out, largely on its own, how to achieve it: what information it needs, what actions to take, in what order, and how to recover when something doesn't go as planned.

Think of the difference between asking a knowledgeable friend a question over coffee, versus asking that same friend to plan and execute your entire vacation — book the flights, reserve the hotel, arrange transportation, and adjust everything if a flight gets cancelled. The first is a single exchange. The second requires the friend to break the goal into steps, take real actions in the world, notice when something changes, and revise the plan. Agentic AI is what happens when we build AI systems capable of that second kind of behavior.

This course is entirely conceptual. You will not write code, install software, or configure a framework. What you will gain is the vocabulary and mental models to understand, and talk knowledgeably about, how agentic systems work, what they can currently do, where they fall short, and how a couple of prominent tools and real products embody these ideas.

Why This Distinction Matters

A conversational LLM — the kind of chatbot you may use daily — has a crucial limitation: it produces one response and then stops. It has no built-in way to check whether its answer was correct, take a follow-up action based on what it just said, or continue working after you stop typing. Every new capability has to come from you, the human, re-prompting it.

Agentic AI removes much of that dependency. By wrapping the language model in a loop — a structure that lets it reason, act, observe the result of that action, and reason again — engineers have built systems that can search the web, run calculations, edit files, book reservations, or coordinate several of these actions in sequence, all in pursuit of a goal that was stated once. This is a significant shift in what AI systems can be used for: not just an oracle you consult, but a delegate you assign work to.

Core Ideas You Will Build

This course develops a small set of interlocking ideas, each covered as its own concept but worth previewing together so you can see how they fit:

Agent — a system with a goal, some ability to make decisions, and some capacity to act autonomously, rather than merely respond.
Goal interpretation and decomposition — how a vague or high-level instruction ("plan my trip to Lisbon") gets turned into a concrete sequence of smaller sub-tasks.
The execution loop — the repeating cycle of reasoning, acting, and observing that lets an agent make progress across many steps instead of just one.
Planning and re-planning — deciding on a sequence of actions in advance, and revising that plan when reality doesn't cooperate.
Tool use — the mechanism by which an agent reaches outside the language model itself to search the web, run code, or call an external service, since an LLM alone cannot directly do any of these things.
Memory — how an agent keeps track of what it has already done and learned so it doesn't repeat itself or lose context across a long task.

The Diagram Behind Almost Every Agentic System

Nearly every agentic AI system built today, regardless of which company or framework produced it, is organized around the same basic loop. Seeing it laid out once here will make every later concept easier to place.

The agent execution loop: a goal enters the cycle of reasoning/planning, acting through a tool, and observing the result; if the goal is not yet met, the loop repeats (dashed arrow back into reasoning) until it is, at which point the agent reports it is done.

What the Course Will Cover, in Order

Starting from this loop, the course will:

Establish precisely what separates an agentic system from a conversational one, and what "autonomy" means in this context.
Walk through each stage of the execution loop in depth — interpreting goals, decomposing tasks, planning, invoking tools, observing outcomes, and re-planning.
Survey the current state of the art: what agentic systems can reliably do today, where they still struggle, and what trends are shaping near-term progress.
Introduce two widely discussed agentic frameworks — describing what problem each is built to solve and how it fits into the broader landscape, without teaching you to use either.
Examine two real, currently available agentic AI applications, tracing how each one's behavior maps onto the general workflow you'll have learned.

By the end, you'll be equipped to follow, and contribute to, informed conversations about agentic AI — what it is, what it isn't yet, and where it seems to be headed — grounded in concepts rather than in any particular product's marketing claims.