Deploying autonomous agents: the practical capabilities of agentic AI 7 minutes

If you’ve been following developments in artificial intelligence, you may have come across a new buzzword: Agentic AI. Agentic AI represents the next big shift, one that will profoundly transform how businesses automate processes and optimize workflows.

This technical deep dive focuses on the current advancements of AI agents across industries, highlighting the main tools used, real-world use cases—like web search, code generation, and optimized SQL queries—and the challenges businesses face when adopting these technologies.

What is Agentic AI?

Put simply, as illustrated in the diagram below, an agentic AI system is fundamentally powered by LLMs (large language models) and built on these core capabilities:

• Memory usage: A well-known strength of LLMs, enabling them to remember increasingly complex sequences of past events. This memory is critical to executing more sophisticated tasks.
• Plan generation: For instance, generating a step-by-step action plan to achieve a goal. A good example would be ChatGPT’s “reason” mode, which generates a plan and reevaluates it after each step.
• Action execution : An LLM can generate SQL queries, code, summaries, analyses, and more. You simply instruct the agent to follow the plan and take action.

A crucial component is tool access. Tools give agents much of their power. For example, if an agent has access to enterprise databases, it can generate a SQL query, execute it, and analyze the results. These tool integrations are essential for realizing the full potential of agentic AI.

Is agentic AI really more advanced than generative AI?

That’s a good question. In fact, agentic AI today is essentially a rather trivial sequence of generative AI prompts. However, when it comes to integrating sophisticated action optimization mechanisms, sky is the limit. Just like with autonomous vehicles or robotics, reinforcement learning and game theory can be used to optimize the sequence of actions.

Key concepts in agentic AI

MCP (Model Context Protocol) format

The Model Context Protocol (MCP) is an open protocol that standardizes the way applications provide context to language models. It acts as a universal connector, enabling AI agents to connect efficiently to various data sources and tools. For example, MCP facilitates the integration of AI agents with databases, web services and other systems, providing a consistent interface for information exchange.

It should be remembered that agentic AI is still new, and this protocol is not yet officially the default. For example, Google recently released its Agent Development Kit (ADK). Roughly speaking, this is a simple, flexible framework for rather simple agents, which could possibly evolve into a protocol depending on its popularity.

The main agentic AI tools

Several tools and frameworks have been developed to facilitate the creation and deployment of AI agents:

• LangChain: A framework for building agents capable of reasoning, interacting with databases and generating text autonomously. It is often used to integrate language models into complex workflows.
• Auto-GPT: An experimental open-source application that demonstrates the ability of language models to perform tasks autonomously, generating code and proactively interacting with APIs.
• BabyAGI: An AI-powered task management system that uses language models and vector databases to create, prioritize and execute tasks based on previous instructions and results.
• LlamaIndex (formerly GPT Index): A framework that enables AI agents to interact efficiently with databases and text corpora by structuring information access in an optimized way.
• CrewAI: Framework for orchestrating AI agents working together collaboratively on complex tasks. It facilitates coordination, communication and role allocation between several autonomous agents.
• PydanticAI: Pydantic extension designed to structure and validate the inputs and outputs of AI models. It simplifies the integration of LLM models with well-defined data types, guaranteeing greater consistency and reliability of responses.
• LangGraph: Framework for creating AI workflows in the form of directed graphs. It facilitates the management of states and interactions between several LLM calls, agents or information retrieval modules.
• OpenAI Agents SDK: Tool developed by OpenAI to design and deploy interactive agents based on OpenAI models. It provides APIs for managing the agents’ memory, tools and conversational capabilities.

Real-world use cases

Web search assisted by AI agents

AI agents are transforming how web searches are performed. Instead of returning a list of links, they analyze multiple sources, compare content, and synthesize insights into coherent reports.

For example, the AutoGLM Rumination agent, developed by Chinese startup Zhipu AI, is capable of performing tasks such as web searches, travel planning and writing research reports. Powered by the GLM-Z1-Air and GLM-4-Air-0414 models, it rivals other agents in terms of performance, while being faster and requiring fewer resources.

Autonomous code generation

AI agents can autonomously generate code, speeding up development and automating repetitive tasks. Tools like Auto-GPT enable users to create, deploy, and operate agents that produce working code based on a defined goal.

In no-code/low-code development, agents are integrated into platforms to generate automated scripts without the need for advanced programming skills, enabling non-technical users to automate complex processes…

Optimized SQL query generation

For companies handling large datasets, query optimization is critical. AI agents can convert natural language queries into highly optimized SQL commands.

An analyst may, for example, describe a business question in natural language, and the AI agent translates this request into an optimized SQL query, taking into account available indexes and best execution practices. Beyond simple generation, an agent can analyze query execution and suggest improvements, such as partitioning a table or adding an index to reduce execution time.

Challenges and constraints of agentic AI

The adoption of agentic AI involves several challenges:

• Supervision and goal alignment: Although agents can perform tasks autonomously, it is essential to establish control mechanisms to avoid unwanted decisions.
• Error management: when an agent generates code or executes a SQL query, it must be able to identify its own errors and correct them without excessive human intervention, requiring advanced self-assessment capabilities.
• System integration: Companies need to adapt their infrastructures to integrate these agents seamlessly, ensuring compatibility with existing systems and interoperability with databases and APIs.
• to internal tools opens the door to security risks. For example, is access secured to prevent the agent from accessing private information? Does the agent have the ability to perform actions that would put system security at risk (e.g., deploying infrastructure, sharing confidential information, etc.)?
• Safety risks Giving an agent access to internal tools opens the door to security risks. For example, is access secured so that the agent doesn’t access private information? Does the agent have the ability to perform actions that would put system security at risk (e.g., deploy infrastructures, share confidential information, etc.)? These questions are essential in any agentic AI project – and echo the best practices discussed in our guide to AI system quality at Moov AIwhere we discuss security, robustness and governance.

Conclusion

Agentic AI is revolutionizing the way organizations automate their processes and exploit their data. Thanks to concepts such as Model Context Protocol and specialized tools such as LangChain and Auto-GPT, these agents are able to perform complex tasks with minimal human intervention.

Whether in web search, code generation or SQL query optimization, concrete use cases demonstrate the immense potential of these technologies. However, their adoption requires in-depth reflection on governance, integration and agent supervision.

The future of agentic AI lies in its ability to adapt in real time to user needs, while guaranteeing enhanced performance and reliability. Companies that know how to exploit these intelligent agents will benefit from a significant strategic advantage in an increasingly automated world.

---

Bibliography

• GitHub. (n.d.). Auto-GPT: An experimental open-source application showcasing the capabilities of the GPT-4 language model. GitHub. https://github.com/Significant-Gravitas/AutoGPT
• LangChain. (n.d.). LangChain: Agents. LangChain Documentation. https://python.langchain.com/docs/tutorials/agents
• LlamaIndex. (n.d.). LlamaIndex Documentation. LlamaIndex. https://docs.llamaindex.ai
• OpenAI. (n.d.). Model Context Protocol (MCP). OpenAI GitHub. https://openai.github.io/openai-agents-python/mcp
• Reuters (2025, March 31). China’s Zhipu AI launches free AI agent, intensifying domestic tech race. Reuters. https://www.reuters.com/technology/artificial-intelligence/chinas-zhipu-ai-launches-free-ai-agent-intensifying-domestic-tech-race-2025-03-31
• Easy with AI. (n.d.). BabyAGI: Autonomous AI Task Management. Easy with AI. https://easywithai.com/ai-agents/babyagi
• Google. (2025, April). Agent Development Kit. https://google.github.io/adk-docs/

Behind the scenes: Our writing process amplified by AI
At Moov AI, we believe in transparency in our use of artificial intelligence. The creation of this article illustrates our vision of AI amplifying human capabilities rather than replacing them.
For the images, we used Midjourney with the following prompt: An abstract illustrationillustration of an silhouette interacting with digital documents, symbolizing the use and integration of artificial intelligence in document management systems. The background is dark blue with subtle waves for depth. There’s a glowing light on one side to highlight user interaction. Bubbly, wavy graphics style.

The images were then modified using Adobe Photoshop (firefly).