Simulate a prompt

A prompt simulation runs a saved prompt version through chat scenarios directly from the Prompt Workbench. There’s no agent definition to create first and nothing to run on your side: the prompt version itself plays the assistant side of the conversation, and Future AGI drives both ends of the chat.

This is the fastest way to see how one prompt version holds up over a multi-turn conversation before it’s wired into an agent or shipped anywhere. If the prompt is meant to sit behind an agent instead, Connect your agent and Create a simulation cover that path; simulating the prompt directly skips both.

Open the Simulation tab

In the Prompt Workbench, open the prompt template you want to test and switch to its Simulation tab.

The tab stays disabled until the template has at least one saved, non-draft version with content in it. Land on it too early and it explains why instead of opening: “You need to submit at least one prompt before running simulations” before anything’s been saved, or “Save your prompt to run simulations” once a draft exists but nothing’s been submitted yet. Save a version first in either case.

Create the simulation

Once the tab opens, its Simulation Runs header lists every simulation already run against this template. Start a new one and a Create Chat Simulation dialog opens with:

• Simulation Name, required
• Prompt Version, the saved version of this template that plays the assistant, shown with Default and Draft chips so you can tell which one you’re picking
• Description, optional
• Select Scenarios, the chat scenarios already in the workspace, with Select All and Deselect All to move fast, plus a Create New Chat Scenario row if none fit yet. That row starts the same process as Create scenarios, just from inside this dialog

Submitting cycles the button through Create Simulation, Creating…, and Starting…, and the run begins on its own from there.

What it runs against

Unlike a simulation built from an agent definition, this one has no agent and no deployment behind it. The prompt version is the thing under test, so Future AGI can call it directly and there’s nothing to connect or drive from your own code. That’s also why it’s chat only: a prompt has no phone number or voice provider attached to it, so a prompt simulation always runs as text, whatever channel the eventual agent will use.

The scenarios work exactly as they do anywhere else: each row plays one conversation, and its persona drives the customer side. See Scenarios for how one is put together.

How results differ from an agent simulation

The run it produces reads like any chat run: Chat Details, not Call Details, throughout, since text is the only mode it ever runs in. The same chat metrics apply, CSAT, token counts, latency, turn count. Voice-only metrics like talk ratio, interruption rate, and words per minute never apply, because a prompt simulation never places a call.

One control an agent-based chat run has is missing here: the Re-run simulation button on the run’s header doesn’t appear for a prompt-sourced run at all. To test a change, create a new simulation instead, most often the same scenarios against a different prompt version, and compare the results of the two runs.

Dive deeper