Modularity Isn't Optional for Agentic Systems

I swapped the embedding model in one of my production systems last week. Took about an hour. No cascading failures, no rewriting half the codebase, no surprises. Just pulled one component out, dropped another in, validated, done.

That’s not luck. That’s modularity.

Agentic systems change fast. Models get replaced. Prompts evolve. Retrieval strategies shift. New capabilities appear every few months that make yesterday’s approach obsolete. If your system isn’t built to accommodate that — if changing one piece means touching everything — you’re going to spend more time fighting your own architecture than building anything useful.

This is component-first thinking. Every piece of the system has a clear boundary, a defined interface, and a specific responsibility. The agent doesn’t care which model is behind the abstraction layer. The retrieval pipeline doesn’t care how the embeddings were generated, as long as the interface contract holds.

We’re heading toward a world of on-demand, composable AI capabilities. Systems that get assembled, used, and reconfigured as needs change. That only works if the pieces are pluggable. If you can pull something out and put something better in without the whole thing falling over.

Modularity isn’t a nice-to-have for agentic systems. It’s the difference between something you can evolve and something you’ll eventually have to throw away.