← Production Insights

#19

Design AI Agents Around Intent, Actions, and State

Don’t design your RAG around endless intents.

Design it around intent + actions + state.

In production AI systems, intent can become a trap.

At first, adding intents feels clean:

  • product_search
  • new_search
  • modify_search

But as the product grows, the intent layer starts becoming a list of every possible variation of the same user goal.

Instead, keep the intent focused on the user's goal.

Then separate that goal from what the system needs to do with the current state.

  • Intent: what is the user trying to accomplish?
  • Action: what should the system do?
  • State: what does the system already know about the current request?

For example:

I need a Ferrari under $50k.

Intent: product_search, action: new_search

Then:

I need it's color to be red.

Intent: product_search, action: modify_search

The intent remains stable.

The actions change the state.

This matters because conversational search is not a sequence of isolated queries.

It is an evolving search state.

Your RAG workflow can remain stable while the user's requirements evolve.

Keep intents stable.

Keep actions flexible.

Let state evolve.