A voice agent answers while the caller is still on the line. Every call to the model is time the caller spends waiting.
On a public voice-agent benchmark with simulated callers, voice agents ran 2x faster with AgentCompile, and made 45% fewer agent calls.
Where the time goes
Most of a voice agent's turn is spent waiting on the model: to understand the request, to choose a tool, to read what came back, and to word the reply. Booking, rescheduling and checking an order are the same jobs every day. When a known job runs compiled, the model isn't called for it.
What stays the same
- Anything new, unclear or unusual goes to your agent's model, unchanged.
- If anything goes wrong, the call goes straight to your model.
- Nothing irreversible happens without the caller's explicit yes.
These numbers are from a public benchmark with simulated callers, not customer data. What you see depends on your own agent's traffic.
How AgentCompile fits voice agents. Read: Voice agents