People sometimes ask how I went from database administration to building AI agents. The short answer is that it was less of a pivot than it sounds. The long answer is this post.
The DBA Years
I spent a significant stretch of my career deep in database platforms. MySQL, PostgreSQL, IBM DB2, both on premises and in the cloud. The work was exactly what you’d expect: query optimization, index tuning, replication setup, backup strategies, capacity planning, and the occasional 2 AM page when something decided to stop working.
What made it formative wasn’t the specific technologies. It was the mental model. When you manage production databases, you develop an instinct for data integrity, for understanding how writes propagate, for thinking about what happens when things fail midway through an operation. You learn to respect transactions. You learn that “eventually consistent” means “eventually someone files a bug report.”
At Vialto Partners, I got to work across a range of data platforms and business contexts. That exposure forced me to think about data not just as rows in a table but as something that drives decisions, enables compliance, and powers the operations that keep a company running.
What Transferred Directly
When I started building AgenticMail, I needed a storage layer that could handle concurrent agent operations, full text search across email bodies, and schema migrations that wouldn’t break a running system. I chose SQLite with WAL (Write Ahead Logging) mode, and that choice came straight from my DBA background.
WAL mode lets multiple readers operate simultaneously while a single writer appends to a log. For an agent coordination system where dozens of agents might be reading email state while one is processing a new message, this is exactly the right concurrency model. I knew this because I’d spent years thinking about read/write patterns in production.
FTS5 (Full Text Search 5) was another decision rooted in database experience. I needed agents to search email content quickly, and I knew from working with PostgreSQL’s full text search that bolting search onto a relational store is almost always better than running a separate search service when your dataset fits in a single node. FTS5 gave me ranked results with BM25 scoring, prefix queries, and boolean operators, all inside SQLite.
The migration system was the third piece. Every DBA has seen what happens when schema changes go wrong in production. The migration system I built tracks versions, verifies content hashes, and refuses to run if someone has tampered with a previously applied migration. That paranoia comes from experience.
What I Had to Learn
The agent coordination side was genuinely new territory. Databases are deterministic: you write a query, you get a result. Agents are probabilistic: you give them a task, and they might interpret it three different ways depending on the prompt, the context window, and apparently the phase of the moon.
I had to learn prompt engineering, tool orchestration, state machine design for agent lifecycles, and how to build guardrails that prevent an AI from doing something catastrophically wrong with someone’s email. None of that was in any DBA handbook.
The biggest mental shift was accepting uncertainty as a design parameter. In database work, you eliminate ambiguity. In agent systems, you manage it. You build fallback paths, confidence thresholds, and human escalation points because you know the system will sometimes get things wrong.
The Connection
Looking back, the through line is systems thinking. A database administrator thinks about the entire lifecycle of data: how it enters the system, how it’s stored, how it’s queried, how it’s backed up, how it’s eventually archived or deleted. An agent system designer thinks about the entire lifecycle of a task: how it’s received, how it’s interpreted, how it’s executed, how errors are handled, how results are verified.
The tools changed. The thinking didn’t.
If you’re a DBA wondering whether your skills translate to the AI space, they do. The industry needs people who understand data at a foundational level, not just people who can call an LLM API. Understanding how data flows, how state is managed, and how systems fail gracefully is exactly what’s missing from most AI agent implementations I’ve seen.
The best agent systems are, at their core, well designed data systems. And that’s something database people already know how to build.