Verification Agent

Modern language programs can write texts, solve tasks, or generate computer code. However, these programs make mistakes on a regular basis, and they sound just as confident doing so as when they give correct answers. A verification agent is a second program that has only one task: to check the result of the first one. It doesn’t invent anything new, but instead looks for contradictions, calculation errors, or fabricated information. The model for this is the division of labor in an editorial office: one person writes the article, a second person proofreads it. The basic idea behind it is simple: checking something is often easier than producing it yourself.

Why Checking Is Easier Than Inventing

Solving a sudoku can take a long time. Checking whether a completed solution is correct takes a minute. Verification agents exploit this imbalance. A comparatively simple checking program can therefore control the output of a very large model without needing to be as powerful itself.

This becomes important everywhere errors are costly. If an AI system independently places orders, answers customer inquiries, or writes program code into a live product, a human is no longer watching every step. This exact lack of oversight is the main reason many companies have so far been hesitant to deploy AI systems. A verification step beforehand significantly reduces the risk.

However, one should not expect too much. A verification agent does not turn an unreliable system into a perfect one. It shifts the error rate downward, often from several percent to a fraction of that. In safety-critical applications, such as in medicine, human oversight remains mandatory nonetheless.

The Process from Answer to Approval

First, the actual model generates an answer. This answer does not go directly to the user, but to the verification agent. It receives the original task along with this, and usually a fixed list of checking criteria. It then issues a verdict: accepted, rejected, or revise.

How the check is performed depends on the task. With program code it’s simple: the agent runs the code and checks whether the test cases pass. The result is then clearly right or wrong. With calculation tasks, it can plug the result back into the original equation. With texts it’s harder, because there is no hard criterion. There, the agent compares claims against a source, such as a stored document, and flags anything that isn’t found there.

If an answer is rejected, it goes back to the first model along with a description of the error. This model tries again, often for several rounds. To prevent this from running endlessly, there is an upper limit on the number of attempts. Each round costs computing time and thus money, which is why the number of iterations is a deliberate trade-off between quality and cost.

Verification Agents in Software and Headlines

They are most visible in programming aids. Tools like GitHub Copilot or similar assistants don’t just write code — they have it tested automatically before suggesting it. Chatbots with source citations work similarly: a checking step verifies whether the cited evidence actually supports the claim.

In the news, the term usually comes up in connection with so-called agent systems. These refer to AI programs that independently carry out multiple work steps in sequence. Companies then advertise that a built-in checking mechanism increases reliability. A related technical term is guardrail, which refers more to fixed rules and prohibitions. A verification agent, by contrast, checks content and makes a case-by-case decision.

A common misconception is to regard the verification agent as an independent authority. Often the same underlying model is behind it as behind the generator, just with different instructions. In that case, both share the same blind spots and overlook the same error. That’s why careful providers use a different model, or an actual program test, as the checker.