What are AI Agents?
AI Agents (or Agentic AI) mark a major shift in how artificial intelligence is applied. A standard Large Language Model interface (like a basic chatbot) is a passive system: it receives text, generates a response, and returns text. An AI Agent is an active software system. It uses an LLM’s reasoning as its central “brain” but pairs it with the ability to call external tools, make API calls, write and run code, and chain multiple actions together to work toward a higher-level goal with limited human intervention.
If an executive tells a standard chatbot, “Find out why Q3 revenue dropped and fix the database error,” the chatbot will reply that it cannot access the database. If the executive gives the same request to an AI Agent connected to the right tools, the Agent can write a SQL query, run it against the Data Lakehouse, read the output, notice that an ingestion pipeline failed, draft a script to fix the pipeline, run it in a controlled environment (with approval), and send the executive a summary. The Agent moves from conversation to action.
The Architecture of Agency
Building an agentic architecture (often with frameworks like LangChain or AutoGen, and with protocols like the Model Context Protocol for tool access) means surrounding the core LLM with programmatic infrastructure.
1. The Tool Registry
The data engineering team gives the Agent a defined registry of Tools. A Tool is usually a Python function or an API endpoint.
The registry might include query_database(), search_web(), read_file(), and execute_code(). For each tool, the registry gives the LLM a JSON schema describing what the tool does and what arguments it requires.
2. The ReAct Loop (Reasoning and Acting)
When given a multi-step task, the Agent works in an iterative loop often described as ReAct (Reason + Act).
- Thought: The LLM analyzes the goal. “I need Q3 revenue. I should use the
query_database()tool.” - Action: The LLM generates the JSON payload to call
query_database(), passing in the SQL it wrote. - Observation: The application runs the function and returns the results to the LLM.
- Thought: The LLM reads the results. “Revenue looks low because a table is missing data. I should investigate the ingestion job.” This loop repeats until the LLM concludes the goal is complete or a configured step limit is reached.
Safety, Governance, and The “Human in the Loop”
Giving AI Agents write access to production databases is a serious security risk. If the Agent makes a mistake, it might take a destructive action such as dropping a Customers table.
Well-designed agentic architectures add “Human-in-the-Loop” (HITL) checkpoints. The Agent can research, write the SQL, and prepare a script on its own, but when it tries to call a tool that changes state (like execute_code()), the orchestration framework pauses. It asks a human engineer: “The Agent wants to run this script. Approve or deny?” The Agent waits until a person explicitly approves, which adds an important layer of protection.
Summary of Technical Value
AI Agents move artificial intelligence from an advisory tool toward a digital worker. By wrapping the reasoning capabilities of Large Language Models in programmatic loops and giving them controlled access to enterprise tools, Agentic AI can automate multi-step data engineering and operational workflows and improve team productivity.
Learn More
To learn more about the Data Lakehouse, read The Lakehouse Built for Everyone by Alex Merced. You can find this and other books by Alex Merced at books.alexmerced.com.