iCallAgent

Appearance

Prompt test ​

A prompt only earns trust once you've watched it handle the calls you didn't design it for. The prompt test is a chat that runs your prompt exactly the way a live call would — same assembly, same model, same settings — so you can push it at the awkward cases before a real caller finds them.

It is deliberately narrower than a test call. A test call exercises the whole stack: speech, voice, latency, tools, telephony. The prompt test isolates one variable — the prompt — so that when something goes wrong you know what caused it.

Where it lives ​

Two ways in, both landing on the same page:

  • Left sidebar → Configure → Prompts → hover a prompt → Test.
  • From inside the prompt wizard, Save & test. The page is keyed by saved prompt, so the wizard saves before opening it.

What it actually runs ​

This is the part that makes the test worth trusting. The page builds the same two things your agent receives on a real call:

text
system prompt  =  the immutable core system prompt
                  +  the prompt under test
                  →  with the agent's own variables substituted

It then sends that to the agent's own configured LLM, at the agent's own temperature and reply-length limit. The only thing substituted is the prompt itself — the prompt under test stands in for whatever prompt the agent currently has, which is what lets you try a prompt before it's attached to anything.

Everything else is read from the agent you select: its variables, its model, its temperature, its greeting. Change agents and all of that changes with it, which is why switching agents clears the transcript — the conversation so far was against a different configuration.

What it does not run ​

Tools and knowledge base retrieval do not fire here. Both live inside the running agent on a real call. Including them would make a disappointing answer ambiguous — the prompt's fault, or a retrieval that missed? — which is exactly what a prompt test exists to rule out.

So expect the agent to say it doesn't know things it would know on a live call, and to be unable to actually book anything. That's the intended behaviour. What you're reading is whether it handles not knowing, and not being able to, the way you want. To exercise tools and retrieval, use a test call.

Running a test ​

  1. Pick an agent under Test against. The badge beside it shows the model, provider and temperature the test will use.
  2. The agent's greeting opens the chat, the way it would open a call.
  3. Type as a caller would and Send (or Ctrl+Enter).
  4. Reset clears the transcript back to the greeting, keeping the same agent and settings.

The last 40 messages are replayed to the model each turn. A long session stays coherent without the context — and the cost — growing without limit.

What to actually test ​

The polite path is the one your prompt already handles. Spend the messages on the rest:

  • Something outside the agent's scope, phrased as though it obviously isn't.
  • "Am I talking to a real person?" — asked plainly, then again more insistently.
  • A caller who's angry, or in a hurry, or who interrupts with a second question.
  • A question whose answer the agent shouldn't guess at — a price, a medical or legal opinion, a promise about a delivery date.
  • Someone who gives you three pieces of information at once, or who trails off mid-sentence.
  • A request to do something the agent cannot do, to see whether it says so plainly or invents a confirmation.

That last one matters more than it sounds. An agent claiming an action succeeded when nothing happened is the failure mode that costs you a customer rather than a call.

Changing the model and settings ​

Click the model badge to open the LLM settings for this test.

ControlDefaultRange
ModelThe agent's configured modelAny active model in the catalog, across all providers
TemperatureThe agent's configured temperatureThe selected model's own min/max
Max reply tokensThe agent's limit, or 800 if it has none64–8,000

The model list here is deliberately wider than the wizard's: you may want to hear how the same prompt sounds on a model you're considering moving the agent to.

Overrides apply to this test chat only. The agent itself is untouched. When any setting differs from the agent's real configuration the badge turns amber and says so — without that it would be easy to tune a great result at temperature 0.2, walk away satisfied, and not notice production is still running at 0.7. Agent defaults restores everything, and switching agents resets the overrides too.

To make a setting real, change it on the agent: left sidebar → Configure → Agents → open the agent → the LLM settings on Agent Overview.

Billing ​

Test replies come out of the same allowance as wizard messages: 50 free per calendar month, then charged to your account balance. See free messages and billing.

Cost is calculated from the model that actually ran — so if you override the model, you're billed at the override's rate, not the agent's.

The first billable message of a session asks for confirmation. Confirming in the wizard also covers testing, and vice versa: it's one acknowledgement per browser session across both.

Next ​