# How to Test an AI Agent with a Synthetic Persona

Try a synthetic persona test for an AI agent, with a copyable prompt, French and English results, and review questions.

- Canonical URL: https://www.harmate.com/en/blog/how-to-test-an-ai-agent-with-a-synthetic-persona
- Author: Harmate Team
- Published: 2026-11-15
- Updated: 2026-11-15T01:54:52.836817+00:00
- Language: en

## Content

# How to Test an AI Agent with a Synthetic Persona

Both instructions explain how to check an order before dispatch. One asks the learner to compare product codes, descriptions and quantities. The other adds a second check of each code and label, drawing on the fictional learner’s own description. To see how an agent uses a profile, run the same scenario with and without it. Then check the source for each adaptation.

## A short test to copy

This condensed prompt lets you run the exercise with a single profile field.

```text
Write a short instruction in English for practising an order check before dispatch. Propose a first activity. Use only the information supplied; identify any profile fields used and what they contribute. Separate established facts from proposals. Say what still needs confirming to schedule the session and to know whether it can be conducted in English.

Scenario: A practice session on checking an order is planned.
Profile: [paste the profile below, or replace it with null]
```

For the profile, the field relevant to this test is copied from the public example:

```json
{
  "skills_use_scenarios": "When there is a huge rush of orders, I remain hyper-methodical to avoid making mistakes. I check every code and every label twice."
}
```

Run the prompt once with `null`, then once with the field. Compare the answers: does the proposed action change, and does the cited field support it? To test another piece of information, change one sentence in the scenario. For example, add that the participant and trainer have confirmed their availability on Tuesday from 10:00 to 10:15. Check whether the answer carries that slot forward without inventing a calendar date or location.

## What six answers showed

The full case used five fields from [Timeo LESCANNE’s public profile](https://www.harmate.com/en/pricing/api#personae-example-title) (revision 46), an order picker who works in person. The five fields cover occupation, work mode, described practice, learning method and languages. The profile lists French at an unspecified level. Whether the learner can do the exercise in English remains open.

| Variant | Observed in the recorded answers |
|---|---|
| No profile | The English answer suggests comparing product code, description and quantity. It says no profile was supplied and leaves scheduling and English suitability open. The French answer likewise proposes checking references and quantities. |
| Profile supplied | Both English and French answers propose checking each code and label twice, linking that adaptation to the field describing this practice. Both leave scheduling and English ability open. |
| Slot added to scenario | Both languages carry Tuesday 10:00–10:15 forward as confirmed. Both propose watching a demonstration before practising. English suitability still needs confirming. |

In the English answer with a profile, the agent attributes the instruction to pause on a mismatch to the same field. The profile says the person checks twice; it does not say how they respond to a discrepancy. Setting an item aside may be a sensible step, but the agent should label it as its proposal, not as a profile habit. Visible input facts let you review that attribution sentence by sentence.

The added slot shows a boundary: the answer carries over the time but still asks for the calendar date and location. Across all six answers, English ability remains open. None treats the mention of French as proof of English proficiency or lack of it.

## How to review your run

For each difference, ask three questions: which profile fact supports it? Which part is the agent’s proposal? What information is still missing? If you add a time slot, separately check which details were carried forward and which remain unresolved. A plausible answer is not evidence of stable behaviour: each variant here was generated only once.

On 15 November 2026 we ran one answer per condition and language in six separate Codex sessions with Luna as the requested model. The results describe those six answers. A fresh run may produce different wording.

This fictional persona provides controlled inputs for testing software. To build a customer profile from field evidence, see [Buyer Persona: Build It from Evidence, Not Assumptions](/en/blog/buyer-persona-build-it-from-evidence-not-assumptions).

To try the exercise with all five fields, [open the public example from the Harmate persona library](https://www.harmate.com/en/pricing/api#personae-example-title) and select **Show the Complete JSON Profile** under Timeo LESCANNE. Copy `profession`, `work_mode`, `skills_use_scenarios`, `learning_and_research_methods` and `languages` into the prompt’s profile block.

If a fact arrives later in the conversation, [try these multi-turn chatbot test cases](https://www.harmate.com/en/blog/chatbot-test-cases-for-multi-turn-conversations). To check the same expectations after a change, [save a versioned regression test case](https://www.harmate.com/en/blog/build-an-ai-agent-regression-test-with-versioned-persona-data).

To prepare your own input data, use the [three-part persona template](https://www.harmate.com/en/blog/synthetic-persona-template-a-minimal-card-for-testing-an-ai-assistant). It records the person’s context, the scenario, and the rules the answer must follow in separate fields.