Agent facing analytics
Agent-facing analytics concepts
What are "agents"?
One can think of AI agents as digital assistants that have evolved beyond simple task execution (or function calling): they can understand context, make decisions, and take meaningful actions toward specific goals. They operate in a "sense-think-act" loop (see ReAct agents), processing various inputs (text, media, data), analyzing situations, and then doing something useful with that information. Most importantly, depending on the application domain, they can theoretically operate at various levels of autonomy, requiring or not human supervision.
The game changer here has been the advent of Large Language Models (LLMs). While we had the notion of AI agents for quite a while, LLMs like the GPT series have given them a massive upgrade in their ability to "understand" and communicate. It's as if they've suddenly become more fluent in "human" aka. able to grasp requests and respond with relevant contextual information drawn from the model's training.
AI agents superpowers: “Tools”
These agents really shine through their access to “tools”. Tools enhance AI agents by giving them abilities to perform tasks. Rather than just being conversational interfaces, they can now get things done whether it’s crunching numbers, searching for information, or managing customer communications. Think of it as the difference between having someone who can describe how to solve a problem and someone who can actually solve it.
For example, ChatGPT is now shipped by default with a search tool. This integration with search providers allows the model to pull current information from the web during conversations. This means it can fact-check responses, access recent events and data, and provide up-to-date information rather than relying solely on its training data.
Tools can also be used to simplify the implementation of Retrieval-Augmented Generation (RAG) pipelines. Instead of relying only on what an AI model learned during training, RAG lets the model pull in relevant information before formulating a response. Here's an example: Using an AI assistant to help with customer support (e.g. Salesforce AgentForce, ServiceNow AI Agents). Without RAG, it would only use its general training to answer questions. But with RAG, when a customer asks about the latest product feature, the system retrieves the most recent documentation, release notes, and historical support tickets before crafting its response. This means that answers are now grounded in the latest information available to the AI model.
Reasoning models
Another development in the AI space, and perhaps one of the most interesting, is the emergence of reasoning models. Systems like OpenAI o1, Anthropic Claude, or DeepSeek-R1 take a more methodical approach by introducing a "thinking" step before responding to a prompt. Instead of generating the answer straightaway, reasoning models use prompting techniques like Chain-of-Thought (CoT) to analyze problems from multiple angles, break them down into steps, and use the tools available to them to gather contextual information when needed.
This represents a shift toward more capable systems that can handle more complex tasks through a combination of reasoning and practical tools. One of the latest examples in this area is the introduction of OpenAI's deep research, an agent that can autonomously conduct complex multi-step research tasks online. It processes and synthesizes information from various sources, including text, images, and PDFs, to generate comprehensive reports within five to thirty minutes, a task that would traditionally take a human several hours.
Real-time analytics for AI agents
Let's take the case of an agentic AI assistant with access to a real-time analytics database containing the company's CRM data. When a user asks about the latest (up-to-the-minute) sales trends, the AI assistant queries the connected data source. It iteratively analyzes the data to identify meaningful patterns and trends, such as month-over-month growth, seasonal variations, or emerging product categories. Finally, it generates a natural language response explaining key findings, often with supporting visualizations. When the main interface is chat-based like in this case, performance matters since these iterative explorations trigger a series of queries that can scan large amounts of data to extract relevant insights.
Some properties make real-time databases especially suitable for such workloads. For example, real-time analytics databases are designed to work with near real-time data, allowing them to process and deliver insights almost immediately as new data arrives. This is crucial for AI agents, as they can require up-to-date information to make (or help make) timely and relevant decisions.
The core analytical capabilities are also important. Real-time analytics databases shine in performing complex aggregations and pattern detection across large datasets. Unlike operational databases focusing primarily on raw data storage or retrieval, these systems are optimized for analyzing vast amounts of information. This makes them particularly well-suited for AI agents that need to uncover trends, detect anomalies, and derive actionable insights.
Real-time analytics databases are also expected to deliver fast performance for interactive querying, essential for chat-based interaction and high-frequency explorative workloads. They ensure consistent performance even with large data volumes and high query concurrency, enabling responsive dialogues and a smoother user experience.
Finally, real-time analytics databases often serve as the ultimate "data sinks" effectively consolidating valuable domain-specific data in a single location. By co-locating essential data across different sources and formats under the same tent, these databases ensure that AI agents have access to a unified view of the domain information, decoupled from operational systems.
These properties already empower real-time databases to play a vital role in serving AI data retrieval use cases at scale (e.g. OpenAI's acquisition of Rockset). They can also enable AI agents to provide fast data-driven responses while offloading the heavy computational work.
It positions the real-time analytics database as a preferred "context provider" for AI agents when it comes to insights.
AI agents as an emerging user persona
A useful way to think about AI agents leveraging real-time analytics databases is to perceive them as a new category of users, or in product manager speak: a user persona.
From the database perspective, we can expect a potentially unlimited number of AI agents, concurrently running a large number of queries on behalf of users, or in autonomy, to perform investigations, refine iterative research and insights, and execute tasks.
Over the years, real-time databases have had the time to adapt to human interactive users, directly connected to the system or via a middleware application layer. Classic personas examples include database administrators, business analysts, data scientists, or software developers building applications on top of the database. The industry has progressively learned their usage patterns and requirements and organically, provided the interfaces, the operators, the UIs, the formats, the clients, and the performance to satisfy their various use cases.
The question now becomes, are we ready to accommodate the AI agent's workloads? What specific features do we need to re-think or create from scratch for these usage patterns?
ClickHouse is rapidly providing answers to some of these questions through a host of features aimed at providing a feature-complete AI experience.
ClickHouse.ai
For more information about features coming soon to ClickHouse Cloud, see ClickHouse.ai.
