Know what your agent actually did.
An agent can act without asking, or claim an action it never took. Voxli records every tool call in every simulated conversation and checks what it did against what it said.
You add a tool, change a schema, or swap the model.
How eagerly an agent calls tools is part of the model's personality. One model cancels the order when the customer only asks what would happen if they canceled it. Another reports the cancellation is done and never calls the tool. Swap the model and both can flip.
The claim, checked against the call record.
Voxli keeps the tool calls next to the transcript. When the agent reports an action with no call behind it, the assertion fails.
Assert the tool, the arguments, and the follow-through.
-
Only reports actions it took
Voxli checks every reply against the call record. A claim without a matching call fails. So does answering 'where is my order?' without ever calling the order lookup.
-
The right chain, the right arguments
Write the check in plain English and Voxli verifies the values: the tool it picked, the order it called them in, and arguments taken from the conversation, even five turns back.
-
Confirmation before it acts
Assert the agent confirms before an irreversible action. We have watched an agent cancel an order because the customer asked what would happen if they did.
Expertise.ai builds conversational AI agents and runs Voxli nightly, with a zero-regression policy on the issues they have already fixed.
Read the storySee your agent's tool calls under test.
We are early stage and founder led. Book a call and talk directly with the people building Voxli.