One agent, many skills: structuring a production AI assistant

We didn't give our AI agent one big brain. We gave it skills.

We built a conversational AI inside an enterprise SaaS platform, where users get work done just by chatting — change a customer's user, create a survey, add questions, reopen customers.

The first version handled everything in one place. It worked fine — while we had 3 or 4 features.

Then the pain started. 😅 Every new feature made that one place heavier. Change one thing, break another. Re-testing the old flows became a nightmare.

Here's what I learned: don't give an AI one giant brain — give it separate "skills."

So we restructured:

→ A user's prompt goes to a classifier that decides which skill is relevant
→ Each skill is self-contained, with its own tools and one responsibility (the "change user" skill only knows how to reassign a user)
→ Common work — like resolving which customer or user the request means — lives in one shared layer, not copy-pasted into every skill
→ Each skill stays in its lane. Not its job? It hands off to another skill.

The payoff: a new feature means adding a new skill, not touching the old ones. Testing became predictable, and changes stopped being scary.

The lesson: if your AI system feels fragile every time you add a feature, the problem usually isn't the AI — it's the structure. Break it into skills.