Dry run of an AI agent: why it's the first step

You've described what you expect from your agent. You've attached your documents, answered its questions, validated its work plan. One step remains before the first real send: the dry run. This simulation shows you exactly what the agent would do, without it actually doing it. No email goes out, no tokens are consumed. You observe, you judge, you correct if necessary.

IWhy test an AI agent before the first send?

An agent that sends an email in your name commits your reputation. If the tone is too casual, if information is missing, if the attachment is wrong, the recipient will not know that a machine drafted it. They will see your signature. The dry run exists to avoid this risk. You verify every detail before anything goes out.

Let's take a concrete example. You are a plumber and you create an agent to respond to quote requests. The agent reads the client's email, identifies the type of work, calculates a price range, drafts a response. During the dry run, you see this response. You notice that the agent proposes a 24-hour intervention deadline when you are alone and overwhelmed. You correct the instruction, you run the test again, you verify that the new deadline appears.

IIWhat is the difference between white testing and draft mode?

A draft mode saves text that you must then send yourself. White testing goes further. The agent simulates the entirety of its work : it reads incoming emails, it decides what to respond, it prepares the attachments, it drafts multiple versions if you asked it to. You see the complete process, not just the final result.

IIIDoes white testing consume tokens?

No. This is one of the major advantages of this step. On ABYS, white testing does not consume tokens. You can run as many tests as necessary to refine the agent's behavior, without touching your monthly reserve. The counter only starts at the first validated real send.

IVHow many times can you run a white test again?

As many times as you deem necessary. Each test shows you what the agent would do with the current instructions. If something does not suit you, you modify the mission, you add a document, you clarify a rule. Then you run it again. The next test incorporates your corrections. This loop continues until you are satisfied.

VWhat should you verify during a white test?

1. 1
   The tone used

   Does the agent use formal language when you want it to? Does the style match your image? A law firm does not write like a pizzeria.

2. 2
   The logic of the responses

   Does the agent answer the right question? Does it suggest the right options? Does it miss essential information?

3. 3
   The attachments

   Does the agent attach the right document? Is the file up to date? Does the format suit the recipient?

4. 4
   Announced deadlines

   Does the agent promise deadlines you can meet? An unrealistic commitment creates more problems than it solves.

5. 5
   Special cases

   What does the agent do when faced with an unusual request? Does it politely refuse? Does it alert you? Does it make up an answer?

VIWhen should you move from white testing to the first real send?

When you have reviewed several tests without finding any significant errors. When the tone suits you, when the logic is sound, when the attachments are correct. At that point, you validate the first send. On ABYS, the first sends remain subject to your manual validation. You have the final say. After three validations without correction, you can switch the agent to automatic validation if you wish.

Without white testingWith white testing

First email sent without review

First email verified and corrected

Errors discovered by the client

Errors corrected before sending

Tokens consumed for tests

No tokens used during testing

Fragile trust in the agent

Trust built step by step

VIIA safety net, not a waste of time

Some find white testing unnecessary. They want the agent to work right away. That is understandable when time is short. But the few minutes spent reviewing a test avoid hours of cleanup. An awkward email sent to an important client can cost far more than a quarter hour of verification.

White testing transforms agent creation into a progressive and controlled. You don't let a machine loose in the wild. You observe it, you adjust it, you validate it. When it starts working for real, you know exactly what it will do.

?Questions we're asked

Does white testing work with all types of agents?

Yes. Whether your agent responds to emails, chases unpaid invoices or prepares quotes, white testing simulates its complete work. You see what it would do in each situation before it actually acts.

Can I do a dry run after already launching the agent?

Yes. If you modify the mission or documents of an already active agent, you can run a white test again to verify the impact of the changes before they apply to the next sends.

What if I am never satisfied with the test?

You continue to adjust the instructions and run new tests. There is no limit. If the result still does not suit you, it may be that the requested task is too complex or poorly defined. Rephrase the mission in simpler terms.

Does white testing also show possible errors?

It shows what the agent would do with the data it has. If an incoming email is ambiguous, you will see how the agent interprets it. That is precisely the point: to spot cases where its logic diverges from yours.