The Database Was the Agent All Along

The AI agent discourse in October 2023 is — predictably — about the agents. Who orchestrates whom, which LLM plays the planner, how many hops until the thing hallucinates itself into a wall. It's all very exciting and almost entirely wrong about where the problem actually lives.

The problem is state. Agents don't fail because the orchestration graph is wrong. They fail because they have nowhere to put things — nowhere persistent, nowhere queryable, nowhere that exists outside the context window that is already half-full of system prompt.

Which is why this Autogen series is worth watching. The architecture isn't novel: a group of agents, each with a role, passing messages back and forth. What's different is that the whole thing is built around Postgres. Not a vector store, not a Redis cache bolted on as an afterthought — a relational database, with tables, with joins, with the boring powerful machinery we've had for forty years sitting underneath everything.

This is the correct answer. It just doesn't feel like the correct answer because we expected the correct answer to be new.

The dream of the autonomous agent is really just a dream of a well-designed application where the application layer happens to be a language model. And well-designed applications have always had the same foundation: a database that knows what happened, keeps it durable, and lets you ask questions about it later.

Agents are the new ORM, in a way — a leaky abstraction over something that was already working fine. The interesting work is in what you put in the schema.

Go watch the series. Ignore the word "agent" in the title. Watch it like it's a database design video. That's what it actually is.