Agentic AI, beneath all the hype, is actually just normal software. You can build agentic systems with a database, standard web tooling and real-time sync.

LangChain, vector databases, instruction routing, specialized memory stores. You'd be forgiven for thinking you need a whole new stack to build agentic systems. However, that isn't actually the case:

Agents are essentially processes that instruct LLMs to make tool calls. Instructing LLMs means sending an instruction to an LLM. Agentic memory is where you store data that those instructions are based on. Retrieval is the ability to query that data. Context engineering retrieves and formats the right information to send in the instruction.

There's obviously a lot of work involved in putting those aspects together to create a working agentic product. However, from an *infra* point of view, there's nothing there that doesn't pattern match to a database, some processes and real-time sync.

For me, as a generalist software engineer, that's really exciting. Because it brings this whole revolution of agentic systems right back into my wheelhouse.