The problem
Every time my real-time voice agent called a tool, it said something like “Let me look that up for you, please hold” first. On a call that should open with “Hi Sarah”, that’s awkward.
What I learned
- The spoken preamble is a platform wrapper around the orchestrator’s plan “thought”. It isn’t generated from your instructions.
- No instruction wording removed it, across several attempts.
- But renaming the tool and rewording its description changes what gets said. The preamble is built from how the tool is described.
- The only way to get no preamble is to run the tool deterministically from a topic rather than letting the orchestrator choose it.
Takeaway
When a generative platform says something you didn’t write, look at the metadata it’s reading (tool names and descriptions) before you fight the prompt.