Model Test

Model test takes a reply the agent has already given and runs that same message on other models, side by side. You see what each candidate would answer and which tools it would call — without touching the real conversation and without executing any action.

It answers one specific question: would this other model handle this message well? It is not a statistical measurement. It is one concrete case, picked by you, run once.

Who sees the option

The option appears for people who can edit agents in the organization. It also depends on the message: it only exists on an agent reply.

The option does not appear:

on your own message, or any message written to the agent;
on a message sent by a human attendant;
in room conversations;
in the public chat or the widget, where the other side is a visitor.

Opening the test

In the conversation, open the action menu on the agent’s reply and choose Test on other models. The Model test window opens.

At the top sits the Original reply: the model that answered, the tools it called, the full text, the credits that turn cost and the date.

If the agent’s prompt changed after that message, a warning says so: This agent’s prompt changed after this message. The original reply came from another version. Read it before comparing — the original reply came from a configuration the test no longer reproduces.

Picking the models

Check up to 3 models among those available in your organization’s plan. When you check a fourth, the screen tells you to uncheck one first.

The list only offers text models. Image generation models do not appear, because they do not produce what the test compares — neither a written reply nor a tool call.

If the organization has no model plan configured yet, or if there is no text model other than the one that already answered, the window says so instead of showing the list.

What each candidate receives

Exactly the same context as the original message, rebuilt at test time: the agent’s prompt, the conversation history up to that message, the connected tools, the conversation summary, memory and the linked knowledge bases.

The history stops at the tested message. The candidate sees nothing that came after it — not even the reply the original agent gave.

Cost

Each selected model is a real call to the provider, and it is charged like a normal turn: one charge per model, with no free allowance. Before running, the window shows Estimate: ≈ N credits and the organization’s Current balance.

An organization with its own key (BYOK) pays the provider directly, and the platform charges 1 fixed credit per model — the estimate reflects that.

Without enough balance the test does not run, and the window asks you to add credits.

Tools: nothing actually runs

This is the feature’s most important rule. Testing three models on one message must not send three emails, create three events or hand the conversation off three times. So:

No write tool executes, and there is no way to turn that on.
If the candidate calls the same tool the original reply called, it gets back the result recorded in that turn — enough for it to keep reasoning.
If it calls a tool the original did not use, or a different one, there is no recorded result to serve as an answer. The test stops there and shows which tool it would have called.

The only exception is the native Search Knowledge Base tool, which only reads. It runs for real, because measuring whether the candidate searches the base when it should is half of what the test exists to answer.

Reading the result

Each model becomes a column, with a status label:

Label	What it means
Same tools	The candidate walked the same tool path as the original reply.
Different tools	It reached the end, but with a different set of tools.
Stopped early	It tried to call a tool with no recorded result, and the test stopped there.
Error	The model did not answer.
Running	The column is still arriving.

Below the label come the candidate’s reply, the tools it called, how many credits that column cost and how many steps it took.

A column that stopped early is still useful information: it shows that the model would pick a different path, and which one.

To try other candidates, use Pick other models in the footer: the comparison leaves the screen and the list comes back, with the original reply and the estimate still there. The columns that were on screen do not come back, and running again charges again.

What the test does not do

It does not affect the real conversation. Nothing the candidates answer is written to the conversation, and the person on the other side sees nothing.
It is not saved. The result lives on screen: while you stay on the same message, closing and reopening shows what has already arrived; leaving discards it. There is no test history.
Switching messages while a test is running cancels it, and the screen warns you when that happens.