Answer caching for AI agents (reuse answers to repeated questions)

Problem
Agents built in Agent Builder make at least one model request for every message, including questions that are identical or very similar to ones already answered. Knowledge-base support bots get many repeat questions, so customers pay credits again and again for the same answer. For high-volume apps, repeat questions can be a large share of credit usage.

Request
Let an agent reuse a previous answer for an identical or very similar question without a new model request, so it uses no credits (or fewer).

Questions this should answer for customers

Can the agent reuse an answer for the same or a very similar question?
How long are cached answers kept?
When does a new model request happen instead?

Suggested behaviour
• Per-agent setting to turn caching on or off.
• Match on exact and semantically similar questions, with an adjustable similarity threshold.
• Configurable time-to-live, for example 24 hours or 7 days.
• The cache clears automatically when the agent’s knowledge base or instructions change, so answers don’t go stale.
• Separate caches when different users have access to different knowledge base content.
• Usage reporting that shows cache hits, so customers can see the credits saved.

Considerations
• Answers that depend on who is asking (for example, account details from a tool call) shouldn’t be shared across users.
• Answers that triggered a tool action shouldn’t be cached.

Impact
Lower and more predictable credit costs for high-volume FAQ use cases, which makes the credit model easier to justify for customers.