Databases for Agents
Luke Wroblewski of Sutter Hill Ventures discusses AgentDB, a database system designed for AI agents, and a paradigm shift in software development where building precedes designing, driven by the power of large language models.
A Paradigm Shift in Application Development
A fundamental shift is occurring in how applications are built, moving towards a simple yet powerful paradigm: an application is increasingly just a URL and a model. This transition mirrors previous platform shifts like the web and mobile, where initial applications seemed less robust but eventually evolved into complex systems.
In this new model, the URL acts as a server endpoint, and the AI model itself becomes the application logic, replacing the traditional “running code” component. AI coding agents can now handle tasks like writing and debugging code. The missing piece has been a database system designed for how these agents operate. When AI models can effectively use a database and render user interfaces from it, the need for developers to directly manipulate code diminishes. The focus shifts to pointing the model at a well-structured data source from which it can generate UIs, content, and interactions.
Introducing AgentDB: A Database for AI Agents
AgentDB was created to solve a core problem: traditional databases were designed for humans, with human-centric interfaces and human-scale operations. AI agents, however, operate at a completely different scale. As noted by Databricks, agents may create thousands of databases per day, and in the future, 99% of databases could be created by agents.
AgentDB is built for this new reality, where “thinking machines,” not humans, are the primary users of the database system.
AgentDB Architecture and Key Features
AgentDB is designed to be an OLTP-style system that removes friction for AI developers and agents. Its key characteristics include:
- Frictionless Creation: A new database can be instantiated instantly with just a unique ID. There are no complex setup forms or configuration decisions.
- File-Based Scalability: Each database is stored as a file (supporting backends like DuckDB and SQLite), allowing it to scale with the file system. This approach scales storage rather than compute.
- Isolation and Portability: Because each database is a file, it is fully isolated. This makes it easily downloadable, deletable, and reusable, mirroring the characteristics of a file system.
- Contextual Templates: When a database is created, it’s paired with a template that provides the AI model with all necessary context. This template includes the database schema, sample queries, migration information, and data formatting rules. This allows an agent to write a correct query “on the first token,” saving significant time, reducing errors, and eliminating the repetitive process of learning the database structure.
Practical Use Cases
The power of AgentDB is evident in its ability to quickly turn simple data into interactive applications:
- CSV to Chat App: Users can upload a CSV file (e.g., credit card statements, a job application tracker) and AgentDB automatically generates a database, a template, and a URL. This URL can be fed to a model like Claude to instantly create a chat application for querying and interacting with that data.
- Agentic CRM: A go-to-market professional combined multiple CSVs (website visitors, GitHub repo stars, contact form submissions) into a single AgentDB database. They then used an agent to enrich this data by searching the web for information on each lead, filling in the details, and organizing the results by priority and region. This creates a highly customized, “agentic AirTable” tailored to a specific workflow.
The Flipped Software Development Lifecycle
The rise of AI coding agents is flipping the traditional software development lifecycle on its head.
- Traditional Model: A significant amount of upfront time was spent on design (writing specs, creating mockups) before the expensive and time-consuming building process began.
- New Model: Developers can now build a functional manifestation of an idea almost instantly. This working prototype is then used to think through design, product integration, and launch decisions. The process becomes Build -> Design, rather than Design -> Build.
This shift requires an adjustment for teams. Design’s role evolves from upfront creation to integrating and creating coherence across rapidly-built features. While this changes the workflow, the core challenge of managing complexity (“Oh crap” moments) simply moves from the engineering phase to the design and integration phase.
Rethinking AI Infrastructure and the Importance of Taste
AgentDB is one piece of a larger puzzle. Much of our existing core infrastructure was built for human interaction and needs to be rethought for an agentic world. Search APIs, for example, are structured around showing a human “10 blue links,” whereas an AI model can consume an entire website for context.
Finally, as agents and automated systems proliferate, the uncommoditized skill will be taste. Simply deploying a technology is not enough. The real work lies in the continuous evaluation and refinement of the system’s output. Using his personal chatbot as an example, Luke Wroblewski emphasizes that he is constantly reviewing user questions, evaluating the model’s responses, and course-correcting the system. This application of taste and perspective is what defines quality and turns a generic tool into a genuinely useful and accurate one.