Agentic AI: From Generative AI to Autonomous AI Agents

AI-103: Developing AI Apps and Agents on Azure

Generative AI has fundamentally changed how humans interact with artificial intelligence. Instead of writing traditional programs for every task, users can simply provide a prompt and let an AI model generate a response.

But what happens when we want AI to do more than just respond?

This is where Agentic AI comes into play.

Agentic AI extends generative AI by allowing AI systems to plan, make decisions, use tools, and execute tasks with a degree of autonomy.

1. Generative AI vs. Agentic AI

The easiest way to understand Agentic AI is to compare it with traditional Generative AI.

Generative AI

Consider a user named Alice. Alice sends a prompt to a Generative AI application:

The AI generates the Python code. The interaction looks roughly like this:

The model generates an answer based on the instructions it receives.

The important point is that Alice is still driving the process. If Alice wants the AI to do something else, she provides another prompt.

2. What Is Agentic AI?

With Agentic AI, the interaction is different. Alice can provide an instruction such as:

Instead of simply returning an answer, an AI agent can potentially:

1. Understand the objective
2. Break the problem into smaller tasks
3. Decide which tools it needs
4. Call those tools
5. Analyze the results
6. Take additional actions
7. Iterate when necessary
8. Produce a final result

The key difference is action.

An agent can therefore operate with a degree of autonomy, within the boundaries defined by its instructions, permissions, tools, and governance.

3. What Makes an AI Agent Different?

An AI agent typically combines several capabilities.

1. Instructions

The agent receives a goal or set of instructions. For example:

The instructions establish the agent's role and boundaries.

2. A Model

The agent needs an underlying AI model to interpret instructions, reason about tasks, and generate responses.

For example, an agent could use a model from providers such as OpenAI, Microsoft, Anthropic, or others available through the platform.

3. Tools

Tools allow an agent to interact with the outside world. For example, an agent could be given access to:

• Web search
• APIs
• Databases
• Knowledge bases
• Code execution
• Microsoft Learn
• Enterprise applications
• File systems
• Other agents

Without tools, an agent is largely limited to what the model can generate from its existing context. With tools, it can retrieve information and perform actions.

4. Memory and Context

Agents can maintain information about an ongoing interaction. This allows them to handle multi-step conversations and workflows rather than treating every request as an isolated prompt.

5. Orchestration

An agent can determine which actions need to happen and in what order.

This orchestration capability is one of the major differences between a simple GenAI application and a more sophisticated agentic system.

4. Real-World Examples of Agentic AI

Agentic AI is particularly useful when a task involves multiple steps.

Autonomous Coding Agents

A traditional GenAI application might receive "Write Python code to calculate the average of a list." It generates the code and stops.

A coding agent could instead:

1. Understand the coding requirement
2. Write the code
3. Execute the code
4. Detect an error
5. Modify the code
6. Run it again
7. Test the result
8. Potentially prepare the final implementation

The agent is therefore participating in the entire workflow, rather than simply generating a code snippet.

Customer Support Agents

A traditional GenAI solution could help a human support representative draft a response:

Support employee --> GenAI --> Suggested response

The employee still needs to process the ticket. An agentic system could potentially:

1. Read the support ticket
2. Identify the customer's problem
3. Search the knowledge base
4. Check the customer's account
5. Determine the appropriate resolution
6. Perform an authorized action
7. Respond to the customer
8. Close the ticket

This creates an end-to-end workflow. Human approval can still be required for sensitive actions.

Research Agents

A research agent could be instructed: "Research the latest developments in a particular technology and prepare a summary."

The agent could:

1. Search the web
2. Retrieve relevant sources
3. Compare information
4. Summarize the findings
5. Cite the sources
6. Produce the final report

This is substantially different from simply asking a chatbot "Tell me about this technology."

Workflow Automation Agents

Agents can also coordinate activities across multiple systems.

An agent can potentially coordinate these steps according to the instructions and permissions it has been given.

5. Single-Agent vs. Multi-Agent Systems

Agentic AI doesn't necessarily mean that there is only one agent. You can also have multiple specialized agents working together.

For example, imagine an e-commerce workflow with:

• Customer Agent - handles customer communication.
• Inventory Agent - checks product availability.
• Pricing Agent - determines pricing and discounts.
• Shipping Agent - determines shipping options.

These agents can collaborate to accomplish a larger objective. The architecture could look like:

This approach can be useful when different tasks require different tools, responsibilities, or specialized instructions.

6. Agentic AI Does Not Mean Unlimited Autonomy

One important concept is that autonomous does not mean uncontrolled. An enterprise agent should operate within defined boundaries.

For example, an agent might be allowed to:

• Search the internet
• Read internal documentation
• Create draft emails
• Query a database

But it might not be allowed to:

• Delete production data
• Transfer money
• Change security policies
• Deploy production infrastructure

without additional authorization.

Therefore, enterprise Agentic AI requires appropriate:

• Identity
• Permissions
• Access control
• Monitoring
• Security
• Governance
• Human approval where appropriate

The more powerful the agent becomes, the more important these controls become.

7. Agentic AI in Microsoft Foundry

Microsoft Foundry provides capabilities for building and deploying AI agents. The Foundry Agent Service can be thought of as an assembly line for AI agents.

The process starts with a business use case. You then select an appropriate model and configure the agent. After that, you can provide:

• Instructions
• Tools
• Knowledge
• Workflows
• Security controls
• Observability
• Governance

The resulting agent can then be deployed for use in applications and business processes.

8. Tools Make Agents Powerful

One of the most important concepts in Agentic AI is tool use. An agent can be given access to different tools depending on what it needs to accomplish.

This means that the model itself doesn't necessarily have to perform every task. Instead, it can decide "I need a tool to accomplish this", then invoke the appropriate tool and use the result.

9. Example: Web Search Agent

Suppose we create a research agent with the instruction:

We can give the agent a web-search tool. Now a user asks: "What are the top trending technologies in 2026?"

The agent can:

The important part is that the agent isn't relying exclusively on information already present in the model's context. It can retrieve current information using a tool.

10. Example: API-Based Agent

Agents can also interact with external APIs. Imagine an activity-planning agent. The user says:

The agent can determine that it needs information from an external activity API. It then:

1. Calls the API
2. Retrieves available activities
3. Filters the results
4. Selects relevant activities
5. Presents them to the user

The API becomes a tool available to the agent. This is an important Agentic AI pattern:

11. MCP and Agentic AI

Another important concept is Model Context Protocol (MCP). MCP provides a standardized way for AI applications and agents to connect to external tools and data sources.

For example, an agent can connect to an MCP server that exposes specific capabilities.

Instead of implementing every integration directly inside the agent, the agent can leverage the tools exposed by the MCP server. This can make agent architectures more modular and extensible.

12. Code Interpreter Agents

Another powerful agent capability is code execution. With a code interpreter, an agent can potentially:

1. Receive a data-analysis request
2. Inspect a dataset
3. Generate Python code
4. Execute the code
5. Analyze the results
6. Create charts
7. Correct errors when necessary
8. Return the findings

For example:

The user doesn't necessarily need to provide the Python code. The agent can generate and execute the code itself in a controlled execution environment.

This makes agents particularly useful for:

• Data analysis
• Mathematical calculations
• Statistical analysis
• Visualization
• File processing
• Code-related tasks

13. The Agentic AI Loop

A useful way to remember how agents operate is the agentic loop:

1. Observe - the agent receives the user's request and available information.
2. Reason / Plan - the agent determines what needs to be done.
3. Act - the agent calls a tool or performs an action.
4. Observe the Result - the agent receives the tool's output.
5. Continue - if additional work is required, it can perform another action.
6. Complete - once the objective has been achieved, the agent returns the final result.

Conceptually:

This loop can occur multiple times before the user receives the final response.

14. Generative AI vs. Agentic AI - Exam Perspective

For certification and technical interviews, remember the fundamental distinction:

The distinction is not simply that agents use LLMs. Most modern agents use generative AI models.

The important distinction is that an agent can use the model as part of a larger system that includes tools, orchestration, memory, actions, and autonomy.

15. Key Takeaways

Agentic AI represents the evolution from simply generating responses to accomplishing tasks.

The most important concepts to remember are:

• Generative AI primarily generates content based on prompts.
• Agentic AI can pursue a goal through multiple steps.
• Agents can operate with partial or significant autonomy.
• Agents can use tools to interact with external systems.
• Tools can include web search, APIs, MCP servers, code interpreters, databases, and knowledge sources.
• Agents can orchestrate multiple actions to accomplish a goal.
• Multi-agent systems allow specialized agents to collaborate.
• Foundry Agent Service provides capabilities for building and deploying agents.
• Enterprise agents require security, permissions, governance, monitoring, and observability.
• Autonomy should always be constrained by clearly defined permissions and business rules.

The simplest way to remember it

That distinction is the foundation for understanding the rest of the AI Agents section in Microsoft Foundry.