In This Handbook

What is Agentic AI?

Agentic AI refers to AI systems that can act autonomously to achieve goals, rather than simply generating responses to user prompts. These systems are called AI agents and are designed to decide what to do next, take actions using tools or interfaces, observe the results, and continue until a task is complete.

Most people encounter AI today through chat interfaces. You ask a question, the model responds, and the interaction ends. Agentic AI behaves differently. Instead of stopping after one response, an AI agent continues working. It reasons about progress, adapts to new information, and takes responsibility for moving a task forward.

This ability to operate over time is what gives agentic AI its name.

Agentic AI meaning

The term agentic comes from the idea of agency, the capacity to act with intent. Applied to AI, it describes systems that are not limited to producing outputs, but are capable of decision-making and execution.

In practical terms, agentic AI includes the ability to:

Interpret a high-level goal
Break that goal into smaller steps
Choose which actions to take
Use tools, APIs, or software environments
Adjust behavior when results differ from expectations

An agentic system does not need to be told every step explicitly. It infers what to do next based on context and feedback.

What Is an AI Agent?

At a fundamental level, an AI agent is a system built to operate with intent. Rather than stopping after generating an output, an agent can evaluate a situation, determine the next best step, select appropriate tools, and execute actions in service of a defined goal. The key distinction is that an agent is responsible not just for producing an answer, but for making progress.

This is why the terms AI agent, AI agents, and intelligent agents in AI are often used interchangeably.

One helpful way to think about agents is to distinguish between two broad modes of behavior.

Reactive agents respond directly to inputs from users or their environment. They are well suited for narrow, well-defined tasks where the required action is clear and repetitive, such as responding to requests or triggering predefined workflows.
Proactive agents go further. They plan ahead, anticipate what might be needed next, and take initiative without requiring continuous prompts. These agents are designed to pursue outcomes, not just respond to events, and they can operate across multiple steps with minimal supervision.

Agentic Cycle

Most agents, regardless of their complexity, operate through a recurring loop:

Perceive: The agent gathers signals from its environment. This may include text input, structured data, images, system events, or responses from APIs.
Reason: The agent interprets what it has observed, evaluates its current state relative to its goal, and decides what action to take next.
Act: The agent carries out that decision by invoking tools, calling external services, or triggering changes in its environment.
Learn: The agent incorporates the outcome of its actions into future behavior. In practice, this rarely involves retraining the underlying model. Instead, learning happens through updated memory, feedback signals, or corrections that the agent can reference in later steps.

This cycle is what allows agents to move beyond single-turn interactions and function as systems that adapt and operate over time.

Examples of agents include:

A customer service AI agent that resolves issues end-to-end
An AI voice agent that handles phone calls
A coding AI agent that edits files and runs tests
A browser AI agent that navigates websites
An AI travel agent that plans and books trips

In all cases, the defining trait is the same: the system is responsible for completing a task, not just responding.

Ongoing AI agents news shows growing experimentation with where autonomy should begin and end, especially around tool access, safety boundaries, and human oversight. These discussions are shaping how agentic AI moves from demos into dependable, real-world systems.

Main Components of an Agent

When discussing AI agents, it’s tempting to focus on visible capabilities like planning or tool usage. But underneath those behaviors lies a small set of core components that every effective agent depends on. These building blocks determine how an agent perceives information, reasons about decisions, takes action, and maintains continuity over time.

At the center of the system is the language model, which provides reasoning and language understanding. Surrounding it are retrieval, tools, and memory, each extending what the agent can know, do, and remember. Together, these components transform a standalone model into a functional agent. Let’s walk through them, starting with the LLM.

Large Language Model (LLM)

The LLM acts as the cognitive core of the agent. It is responsible for interpreting inputs, understanding intent, reasoning about goals, and generating plans or responses. Whether the agent is summarizing information, breaking down a task, or deciding which action to take next, the LLM is driving those decisions.

Several factors influence how effective the LLM is in an agentic system:

Its ability to follow instructions and remain coherent across multi-step or multi-turn interactions
The size of its context window, which limits how much information it can consider at once
Domain adaptation through prompting or fine-tuning
Practical constraints such as latency, compute requirements, and cost

On its own, the LLM can reason and communicate—but it cannot access new information or act in the world. That’s where the other components come in.

Retrieval

Retrieval extends the agent’s knowledge beyond what is embedded in the model’s training data. Without retrieval, an agent is restricted to static knowledge. With it, the agent can access current, domain-specific, or proprietary information as needed.

Common retrieval sources include:

Document collections or knowledge bases searched via embeddings or keywords
Live web search for up-to-date information
Internal enterprise systems or structured databases

Many systems use retrieval-augmented generation (RAG), where relevant information is first retrieved and then injected into the LLM’s context. This allows agents to ground their reasoning in external sources rather than relying on assumptions or outdated data.

Tools

While retrieval helps agents know more, tools allow them to do more.

Tools are external functions or systems that an agent can invoke to carry out actions that go beyond text generation. They are what enable agents to affect their environment and complete real tasks.

Typical tools include:

APIs for services like search, weather, or scheduling
Database queries and updates
Code execution environments
File operations or system commands
Integrations for email, web scraping, or workflow automation

When tools are used effectively, agents move beyond conversation. They can fetch live data, perform transactions, update systems, and orchestrate workflows.

Memory

Memory gives an agent continuity across time. Instead of treating every interaction as a fresh start, memory allows agents to retain context, learn from outcomes, and adapt future behavior.

Memory in agentic systems typically includes:

Short-term memory, such as recent interactions or intermediate reasoning steps within a session
Long-term memory, including user preferences, past tasks, outcomes, and feedback
Storage mechanisms, such as vector databases, semantic indexes, graph structures, or structured records
Retrieval logic, which determines what past information is relevant to the current situation
Maintenance strategies, including updating, weighting, or pruning memories to keep the system efficient

Importantly, this form of learning does not usually involve retraining the model itself. Instead, the agent improves by managing what it remembers and how it uses that information in future decisions.

Generative AI vs. Agentic AI

Many modern agentic systems are built on top of Generative AI models, but the two are not the same.

Generative AI focuses on producing content: text, code, images, or summaries based on a prompt. Once the output is generated, the model’s role is complete.

Agentic AI uses generative models differently. Instead of asking for a final answer, the system uses the model to:

Reason about what should happen next
Decide which action to take
Interpret the results of previous actions

In other words, generative AI produces responses, while agentic AI produces progress.

This distinction becomes important as tasks grow more complex. Writing an email may only require generative AI. Investigating an issue, coordinating multiple steps, or operating across systems usually requires an agentic approach.

Agentic AI vs. LLMs

After distinguishing agentic AI from generative AI, it’s worth clarifying how agentic AI relates to large language models (LLMs). The two are often discussed together, but they play very different roles.

An LLM is a model.
Agentic AI is a way of using models.

This difference explains much of the confusion around AI agents.

Building agentic AI applications with a problem-first approach

Once you understand the difference between generative AI and agentic AI, the next question is not which tools to use, but what problem you are trying to solve.

A common mistake when building agentic AI applications is starting with the agent itself, choosing a framework, wiring up tools, and then looking for something useful for the system to do. This often leads to over-engineered agents that are impressive in demos but unreliable in real workflows.

A problem-first approach reverses that order.

Instead of asking “How do we build an AI agent?”, you start by asking:

What task is currently hard to complete?
Where do humans get stuck coordinating steps?
Which decisions are repetitive, time-consuming, or error-prone?
What does success look like from start to finish?

Agentic AI works best when it is applied to messy, multi-step problems, not clean, one-shot tasks.

Manus AI agent is an autonomous AI agent designed to carry out complex tasks independently: like research, data analysis, coding, and web automation—without continuous human guidance, running workflows in the background and delivering results directly.

Clarifying a common notion: “AI agents take control of the computer”

At this point, it’s useful to clear up a common misunderstanding about agentic AI.

When people hear about AI agents acting autonomously, they often assume that an AI agent takes control of the computer: clicking buttons, typing commands, or operating systems without restriction. This idea shows up frequently in discussions about agentic AI, and it’s understandable why it raises concern.

In practice, this is not how most agentic AI systems are designed.

An AI agent does not inherently “take control” of anything. It can only act through explicitly provided interfaces. If an agent is able to interact with a browser, a terminal, or an application, it is because those capabilities were intentionally exposed by the developer.

This distinction matters.

Key Properties of Agentic AI

Let’s discuss the properties of Agentic AI.

1. Goal-Driven, Not Prompt-Driven

In a non-agentic setup, the prompt defines the entire task. If the task changes or fails halfway through, the system has no built-in mechanism to recover.

Agentic AI systems, by contrast, are anchored to goals, not individual prompts. A user might specify a high-level objective such as:

“Analyze last quarter’s sales data and prepare a summary for leadership.”

The agent then determines how to achieve that goal: which data to fetch, how to analyze it, what format the summary should take, and whether additional clarification is needed.

The prompt becomes a starting signal, not the full specification.

2. Planning and Decomposition

Agentic systems can break down complex tasks into smaller steps. This is critical for real-world problems, which are rarely solvable in a single inference call.

For example, given a goal like “Set up monitoring for this service”, an agent might:

Identify what service is being referenced
Check existing monitoring configurations
Decide which metrics are missing
Provision dashboards or alerts
Validate that alerts trigger correctly

Each step may involve different tools and reasoning modes. The agent coordinates these steps without requiring the user to micromanage them.

3. Action Through Tools

Agency requires the ability to do things, not just talk about them.

Agentic AI systems are typically integrated with tools such as:

APIs and SDKs
Databases and data warehouses
File systems
Code execution environments
Internal services

The agent decides when to call a tool, which parameters to use, and how to interpret the response. This is a major shift from traditional automation, where the sequence of actions is hard-coded by developers.

4. Feedback and Adaptation

A defining feature of agentic AI is that it can observe outcomes and adjust behavior.

If an action fails, returns incomplete data, or produces unexpected results, the agent can:

Retry with different parameters
Choose an alternative tool
Revise its plan
Ask for clarification

This feedback loop makes agentic systems more robust in dynamic environments, where assumptions frequently break.

5. Persistence Over Time

Most AI interactions are stateless. Once the response is generated, the system forgets the interaction.

Agentic AI systems often maintain persistent state:

What has already been tried
What information has been gathered
What decisions were made earlier
What remains to be done

This allows agents to work on tasks that span minutes, hours, or even days, something that single-prompt models cannot do reliably.

What Agentic AI Is Not

It’s important to clarify what agentic AI does not mean.

It does not imply consciousness or human-like intent.
It does not mean the system is uncontrolled or unsafe by default.
It does not replace deterministic workflows where strict guarantees are required.

Agentic AI is best understood as a software architecture pattern, not a claim about intelligence or autonomy in the human sense.

Why Agentic AI Matters

As systems grow more complex, the cost of manual orchestration increases. Humans become bottlenecks for:

Sequencing tasks
Handling edge cases
Monitoring long-running processes

Agentic AI shifts this burden onto software. Instead of automating individual steps, it automates decision-making between steps.

This is why agentic AI is increasingly used in areas like:

DevOps and infrastructure automation
Data analysis and reporting
Customer support workflows
Research and knowledge synthesis
Internal productivity tools

In system design terms, agentic AI represents a move from stateless services to stateful, goal-oriented systems, with all the tradeoffs that come with that shift.

Agentic AI Use Cases

Agentic AI is already being used to automate and improve workflows that involve multiple decisions and actions over time. Rather than just generating outputs, AI agents actively drive progress and work across systems to complete tasks autonomously.

Across industries, common Agentic AI use cases include:

Enterprise automation: AI agents can orchestrate multi-step work across business systems, such as IT service management, HR onboarding, and finance workflows, reducing manual handoffs and accelerating resolution times.
Customer support and service: Intelligent agents handle support tickets, triage incidents, and even take action across tools (like resetting access or updating records), improving response times and consistency.
Sales and revenue tasks: AI agents can automate lead qualification, follow-up activities, and customer outreach, freeing sales teams to focus on high-value engagements.
Supply chain and logistics: Agentic AI helps monitor inventory, optimize delivery routes, and adapt plans in real time based on demand changes or disruptions, boosting resilience and efficiency.
Healthcare operations: In clinical and administrative settings, AI agents streamline scheduling, manage patient data, and support decision-making by synthesizing information from multiple sources.
Retail and e-commerce: Personalized shopping assistants and autonomous bots can track customer preferences, recommend products, and manage orders on behalf of shoppers.
Finance and risk management: Agentic systems analyze financial data continuously, manage portfolios, monitor for irregularities, and adapt strategies based on real-time conditions.
Security operations: Security teams are using AI agents to triage alerts, evaluate threats, and automate initial investigation steps, helping reduce burnout and speed up response.

These applications illustrate how agentic AI transitions AI from reactive assistance to proactive execution, handling sequences of tasks that traditionally required human oversight.

Journey forward

With the foundations of agentic AI established, the natural next step is to build an AI agent. Before introducing advanced retrieval or orchestration patterns, it’s important to understand how an agent is created, how it reasons, and how it takes action in a controlled way. The journey therefore begins with hands-on agent construction using frameworks such as LangGraph and CrewAI, which introduce core ideas like state, control flow, and multi-agent collaboration. Once these fundamentals are in place, the focus expands to Agentic System Design and Agentic Design Patterns, providing reusable structures for building reliable and scalable agentic applications. With a solid agent foundation, the course then explores Agentic RAG and Agentic Graph RAG, where agents actively decide how to retrieve and reason over knowledge. Finally, the journey moves into agentic protocols: including MCP and A2A, which define how agents communicate, coordinate, and operate across tools, services, and environments.

Previous Lesson Next Lesson

Fahim Ul Haq

I’m Fahim, a software engineer specializing in generative AI, system design, and responsible AI adoption.

View Profile

Generative AI