Can one AI agent check another agent's work?

A question from a training session

Recently DATONG AI ran a training session at a company on moving from using AI in a chat window to letting an agent do the work directly. After the hands-on part, a participant from the sales team asked: can a second agent check what the first one did?

It is the right question. In day-to-day work, whatever an agent delivers has to be checked, by a person or by a fixed mechanism. What matters is how.

Why not just let one AI review the other?

When one agent reads another agent's views, strategy, code or judgement, one of two things tends to happen. The reviewer nitpicks: it always finds something, even when there is no real problem. Or the agent under review adds more and more safeguards to protect itself, and its output becomes more and more cautious.

On top of that, some mistakes are caught instantly by a simple rule, such as whether a number or a format is right. Handing that to an expensive model to “think about” is a waste. A thermometer tells you whether someone has a fever; you don't need a panel of three specialists. In our own projects we have found that letting agents review each other mostly just burns tokens.

Example: emails to customers

The participant was thinking about customer emails. In the training we break it down together:

• What rules can check: Are the customer's name and company right? Do prices and delivery dates match the quote? Is the attachment there? In German, is the formal “Sie” used where it should be? These become check rules that run on every email automatically, with no model needed to judge.
• What needs business judgement: Is the tone right? Does it sound like our company? Will the customer understand the wording? Does it promise too much? Only people who write to customers every day can judge this. So participants take emails they wrote themselves and consider good, and turn “good” into a few points that can be ticked off.
• Then the agent writes for a set of real situations, and the results are checked point by point. Where problems keep recurring, the agent's instructions are adjusted.

A second AI can help here, but as someone ticking off the sales team's checklist, not as a free-ranging critic. For more demanding roles and tasks, testing is done in layers: format and facts first, then business judgement, and finally the effect in the real workflow.

Why the people doing the work set the standard

Whether an agent is too cautious, or has drifted away from the actual business goal, can only be judged by the people who do the work every day. Outside consultants, IT, even management do not necessarily know better what a good follow-up email looks like. The logic is not hard to learn; many people simply don't think of it at first because they have never programmed and have little experience with AI. That is why, in DATONG AI's training, participants work out this step themselves: what to test and how.

When it is not worth it

For one-off work that never leaves the company, such as your own meeting notes, a quick look is enough. A test system pays off for work that repeats daily or weekly and where a mistake affects customers or money.