In the rapidly evolving world of artificial intelligence, a new player has emerged, capturing the attention of both technologists and business leaders: DeepSeek. This Chinese startup has developed an AI model, DeepSeek-R1, that rivals industry giants like OpenAI's ChatGPT, achieving comparable performance at a fraction of the cost. The impact was immediate and profound: U.S. tech stocks, particularly in the AI sector, saw a significant downturn, with the Nasdaq experiencing a $1 trillion loss in value. Companies such as Nvidia, Microsoft, and Google faced notable declines in their stock prices.
DeepSeek's accomplishment has ignited discussions about the next frontier in AI: Agentic AI. While the concept of agents – systems capable of autonomous actions – has been well-established in fields like robotics and software development, the term Agentic AI has recently gained prominence. It represents a new class of AI systems that combine reasoning, adaptability, and the ability to act with minimal human intervention.
By now, you might be thinking, “Wait, haven’t we seen agent-based terms before?” You’re absolutely right! Let’s take a moment to summarize all the “agentic” concepts humanity has explored so far:
Agentic: Refers to the ability to make independent decisions in pursuit of a goal. It’s the core idea behind systems that are not just reactive but proactive, working towards predefined objectives.
Agentic AI: Builds on this concept by leveraging advanced reasoning and iterative planning to solve complex, multistep problems autonomously. These systems don’t just analyse data; they act on it, adapting their behaviour in dynamic environments.
Autonomous Agents: These are practical implementations of Agentic AI. An autonomous agent is a self-contained software entity that perceives its environment, makes decisions based on that perception, and executes actions to achieve its goals. Autonomous agents, powered by Agentic AI, are the building blocks of this next generation of intelligent systems.
Unlike traditional AI applications, which are often limited to single-use cases, Agentic AI opens the door to multipurpose, goal-driven systems capable of operating independently across industries. Whether it’s managing supply chains, automating customer interactions, or optimizing healthcare workflows, these systems are transforming how businesses think about efficiency and innovation.
Real-World Examples of Agentic AI in Action
To understand the transformative power of Agentic AI, let’s explore how foundational models with agentic capabilities are already making an impact across industries:
DeepSeek-R1 – Reshaping Competitive AI
DeepSeek-R1 exemplifies a foundational model with agentic traits, utilizing reasoning and advanced task-solving capabilities. Its application as an AI assistant goes beyond simple responses, autonomously generating insights and performing tasks like coding and document analysis. By democratizing access through its open-source platform, DeepSeek has sparked a global conversation about the accessibility and affordability of cutting-edge AI systems.
OpenAI’s “Operator” Tool – AI with Hands-on Functionality
OpenAI’s recently unveiled "Operator" tool pushes the boundaries of autonomy by enabling models to interact directly with web interfaces. For example, imagine a customer service agent capable of navigating complex databases, filling out forms, or even booking appointments without human oversight. By bridging natural language processing with actionable interface interaction, "Operator" is bringing businesses closer to full AI-driven automation.
Healthcare Innovation in Switzerland
In Switzerland, the healthcare sector is increasingly integrating Agentic AI to enhance patient care and operational efficiency. For instance, Swiss hospitals are utilizing AI-driven systems to analyse vast amounts of medical data, streamline clinical workflows, and offer more personalized treatment plans for patients. This approach not only improves diagnostic accuracy but also allows healthcare professionals to focus more on patient-centered care.
Automated Price Management in Procurement
DiPriMa, a Swiss startup, has developed a digital platform that automates price management processes for procurement teams. Utilizing AI, DiPriMa captures contract data, refines price adjustment formulas, and executes automated pricing solutions. This approach reduces manual tasks, ensures compliance, and allows professionals to focus on strategic decision-making.
The Secret Sauce
Foundational models like DeepSeek-R1 or OpenAI’s ChatGPT-o1 are nothing short of amazing. They excel at reasoning, language generation, and even solving complex problems. But here’s the catch – they don’t naturally interact with the world. Imagine having a genius locked in a library, brilliant at answering questions, but unable to step outside to actually act on their knowledge. This is where tools come in, bridging the gap between thinking and doing.
Tools transform models into agents. They give these systems the ability to fetch live data, interact with APIs, or even take meaningful actions in the real world. Whether it’s booking flights, managing databases, or retrieving documents from external sources, tools are the enablers that turn a static model into a dynamic agent.
What’s fascinating is how tools interact within the agent’s architecture. For instance, an agent doesn’t just know the weather—it calls a weather API to retrieve live data and then refines its actions based on that information. Think of an agent planning a trip: it gathers real-time flight information, cross-references hotel availability, and dynamically suggests the best itinerary. All of this happens because tools are seamlessly woven into the system, guiding the agent from reasoning to execution.
This entire process is orchestrated through something called the orchestration layer, the true brain of an agent. It’s not a single task; it’s a loop. The agent observes its environment, reasons through its options, takes action, and then adjusts its strategy based on results. It’s like a chef constantly tasting and adjusting their dish until it’s perfect. Frameworks like ReAct (Reasoning and Action) or Tree-of-Thought provide the scaffolding for this loop, allowing agents to refine their approach in real-time.
What makes this exciting is the sheer adaptability. Tools allow agents to connect with real-world data sources, extend their capabilities with APIs, and even integrate client-side control for tasks that require additional security or customization. This isn’t just theoretical—it’s the backbone of how Agentic AI is already transforming industries. Healthcare agents use tools to schedule appointments and monitor patient data. Logistics systems optimize supply chains by adapting to weather disruptions in real-time. These aren’t static models—they’re decision-makers in action.
By equipping agents with the right tools, developers unlock a world of possibilities. It’s not just about answering questions or generating text anymore; it’s about taking meaningful, real-world action. And as tools become more advanced, the line between thought and action continues to blur, opening up a new era for AI.
Expertise Needed to Build Agentic AI
Building Agentic AI isn’t just about choosing the right tools or models—it’s about orchestrating a symphony of skills, technologies, and design principles. This field sits at the intersection of AI research, software engineering, and real-world problem-solving, making expertise as diverse as the applications themselves.
To start, understanding foundational models is essential. These models, like ChatGPT-o1 or DeepSeek-R1, are the core decision-makers of any agentic system. But deploying them effectively means knowing more than just how to query them. Developers need to understand how to fine-tune these models, integrate them with external tools, and optimize them for specific tasks. The ability to work with reasoning frameworks like ReAct or Chain-of-Thought, which enhance the decision-making capabilities of models, is a valuable skill.
Next comes tool integration, which bridges the gap between static models and dynamic environments. This requires a working knowledge of APIs, data pipelines, and extensions that allow agents to interact with external systems. Developers must think strategically about how agents will access and use real-time information—whether it’s through a database, a live API, or a custom function.
Equally important is building the orchestration layer, the core of an agent’s cognitive architecture. This layer handles the complex cycles of observation, reasoning, and action. Engineers working in this space must design systems that manage state, maintain memory, and execute multi-step plans. It’s not just about coding—it’s about creating a structure that allows agents to adapt and iterate intelligently.
Finally, there’s the need for domain expertise. Whether the agent is optimizing supply chains, enhancing customer service, or improving medical outcomes, developers must tailor the system to the specific needs of the field. This often involves close collaboration with subject-matter experts to ensure the agent’s actions align with industry standards and goals.
Building Agentic AI is not just about technical prowess—it’s a blend of creativity, problem-solving, and deep technical skill. And as the technology evolves, so too will the expertise required to harness its full potential. It’s a challenge, yes, but one that brings with it immense opportunity.
Example Tools for Agentic AI
When discussing tools in the context of Agentic AI, we rely on the definitions provided in the Agents whitepaper. In this framework, Extensions, Functions, and Data Stores are the primary tool types that enable agents to interact with the world, whether by connecting to APIs, processing structured outputs, or accessing real-time data. While these names are specific to Google Cloud’s Vertex AI, other platforms may use different terminology to describe similar capabilities. The emphasis here is on the actions an agent can perform and the location of their execution, rather than the specific names.
Extensions: Connecting Agents to APIs
Extensions act as a bridge between an agent and external APIs. They enable agents to perform real-time API calls, dynamically selecting the appropriate extension based on the task at hand. This allows agents to adapt and take meaningful actions with minimal pre-configuration.
Here’s an example of creating an Extension for the SwissIndex, a public database of Swiss business registrations. This extension allows an agent to query the index for specific company information:
import vertexai
PROJECT_ID = "YOUR_PROJECT_ID"
REGION = "us-central1"
vertexai.init(project=PROJECT_ID, location=REGION)
zefix_extension = Extension.create(
manifest={
"name": "zefix",
"description": "Access to the Swiss central business name index",
"api_spec": {
"openApiYaml": ZEFIX_OPENAPI,
},
"auth_config": {
"auth_type": "HTTP_BASIC_AUTH",
"httpBasicAuthConfig": {
"credentialSecret": ZEFIX_SECRET,
},
},
}
)
response = zefix_extension.query("Search for Renuo")
print(response.steps[-1].parts[-1].text)
This example shows how to configure an extension with authentication and API specifications, enabling the agent to interact directly with SwissIndex data.
Functions: Enabling Client-Side Execution
Functions give developers greater control over API execution by shifting it to the client side. While the agent provides the parameters or data needed for the API call, the actual execution happens outside the agent, ensuring secure and customized operation.
Here’s an example of listing available Vertex AI extensions using a Function:
from vertexai.preview.extensions import Extension
def list_extensions():
"""Provides a list of Vertex AI extensions available to the user."""
extensions_list = Extension.list()
return [
(e.display_name, e.resource_name, e.create_time)
for e in extensions_list
]
model = GenerativeModel("gemini-1.5-flash-001")
user_prompt_content = Content(
role="user",
parts=[
Part.from_text("Which Vetex AI extension do I have access to ?"),
],
)
support_tool = Tool(
function_declarations=[list_extensions_function],
)
response = model.generate_content(
user_prompt_content,
generation_config=GenerationConfig(temperature=0),
tools=[support_tool],
)
for function_call in response.candidates[0].function_calls:
if function_call.name == "list_extensions":
api_response = list_extensions()
response = model.generate_content(
[
user_prompt_content,
response.candidates[0].content,
Content(
parts=[
Part.from_function_response(
name=function_call.name,
response={"content": api_response},
),
],
),
],
tools=[support_tool],
)
print(response.text)
The agent can use this function to dynamically generate output while leaving the actual API interaction to the developer’s control.
Data Stores: Extending Knowledge with Real-Time Data
Data Stores allow agents to access real-time or structured data from external sources, such as databases or documents. By integrating a vector database, agents can perform semantic searches and retrieve contextually relevant information.
For example, consider a legal assistant agent retrieving data from a collection of regulatory documents stored in a vector database. The agent uses a Retrieval-Augmented Generation (RAG) process to query the database and integrate the results into its reasoning loop.
Comparison: Extensions, Functions, and Data Stores
While the specific names of these tools—Extensions, Functions, and Data Stores—are based on the Agents whitepaper, their core functionality remains universal across platforms.
Extensions execute actions directly within the agent, enabling seamless interaction with external APIs.
Functions offload execution to the client side, offering more control and flexibility.
Data Stores allow agents to extend their knowledge by accessing dynamic, external datasets.
By combining these tools, developers can design agents that are both powerful and adaptable, capable of reasoning and acting in a wide range of scenarios.
This triad of tools—referred to here as Extensions, Functions, and Data Stores—forms the backbone of any robust Agentic AI architecture, regardless of how different platforms might label them.
The Promise of Agentic AI
When DeepSeek disrupted the AI landscape with its groundbreaking model, the world took notice—not just because of the technical achievement, but because it marked the beginning of a broader shift in how AI systems operate and influence the world. This shift isn’t merely about better models; it’s about smarter systems that can “reason”, act, and adapt in ways that were previously unimaginable.
Agentic AI builds on this promise, transforming the core concept of “agentic” decision-making into a reality that spans industries. What started as an idea—AI systems capable of independent, goal-oriented actions—has matured into autonomous agents that are reshaping how businesses and individuals tackle complex challenges. These agents, powered by tools like Extensions, Functions, and Data Stores, move beyond static intelligence to dynamic, proactive problem-solving.
The journey from Agentic AI as a concept to its practical implementation in autonomous agents demonstrates how technology can evolve from theory to game-changing solutions. DeepSeek-R1 and similar innovations are not isolated achievements; they’re signposts pointing to a future where AI systems don’t just analyze the world—they engage with it, adapt to it, and improve it.
As we’ve explored, the tools and architectures enabling this transformation are robust, yet their possibilities are just beginning to be understood. Whether it’s the dynamic flexibility of Extensions, the controlled precision of Functions, or the knowledge-expanding potential of Data Stores, each component plays a vital role in defining the landscape of Agentic AI.
But as with any powerful technology, the responsibility to guide it lies with us. Designing agents that are not only effective but also ethical, transparent, and aligned with human values is the challenge and opportunity of our time.
The next frontier of artificial intelligence is here, defined not by isolated feats but by integrated, agentic systems that reshape how we solve problems, innovate, and collaborate. Just as DeepSeek’s impact marked a turning point, so too does the rise of Agentic AI signal the dawn of a new era—one where intelligence is not just smart but purposeful, not just reactive but proactive.
The question isn’t whether Agentic AI will transform our world—it’s how we’ll harness it to create a future worth building. The journey has just begun. How will you be part of it?
More information here.
