In software development, changing something to make it better always opens up the possibility of breaking something else. Even a small adjustment can change the behavior of a program, so proposing a change and proving that it works are two different things. Now let’s shift that dilemma to artificial intelligence agents capable of interfering with the code and instructions they use to solve tasks. They can try to perfect their own tools, yes, but each attempt poses the same challenge: when a change is worth it and when it might introduce a problem that didn’t exist before.
Challenge to improve without failing. A team of researchers from the University of Virginia, Microsoft Research, and Tsinghua University has developed a way to give these agents one more chance before testing their changes. The idea, presented in a study published on arXiv, is that a system tries to predict that a change might worsen the outcome and, when it detects that risk, allows it to be reviewed. We are not talking about retraining the artificial intelligence, but improving the programs and instructions used to make it work. The goal is not to eliminate all errors, but to better guide improvement efforts.
This is how agents are changed. Let’s imagine two artificial intelligence agents working together. The first is dedicated to solving tasks, while the second looks for ways to improve its operation, changing the code and instructions used by both. The latter may modify its own software. This is how HyperAgents works, the system from which research is done and which stores results from previous versions to guide future changes. This is not the first attempt to achieve something similar: in 2025, Sakana presented the AI Darwin Godel Machine, a project that has already explored how an agent can improve by changing its own code.
Guess a problem before the test. To determine when it is appropriate to review a change, researchers incorporate the Improvement Dynamics Model (IDM). This component analyzes the proposed change and uses accumulated experience to estimate its likely consequences. You can distinguish, for example, a task that will end with an incorrect answer and one that will not be completed. If you expect worse, allow additional review before evaluating the resulting version. During a test, some instructions were left outside the corresponding quotes and caused a syntax error The agent detected the problem, but ended its calls to the tool before performing the repair. The forecaster speculated that the tasks would probably be left incomplete and another review would be requested. After modifying the code, the new version was able to solve 11 tasks. The system evaluates the final even if the last prediction is unfavorable.
Tests show improvement. The researchers didn’t stop at the quotation mark example: they tested the agents with 166 programming tasks that they didn’t use during the improvement process or to select versions. To consider a task solved, the system had to propose a code change that passed the corresponding test. The initial agent achieved a success rate of approximately 16%, while selected versions of HyperAgent, excluding the new prediction, achieved a median of approximately 22%. With prediction methods, this number reaches about 30%. The results, obtained from five runs per configuration, indicate an improvement over the test conditions.
A prophecy that learns. There is a second capability that deserves attention: the system responsible for anticipating outcomes can also be improved. Every five rounds, the agent responsible for the change proposes a change to its action, which is incorporated only if the next check passes. The researchers also wanted to know whether this experience can be used in various tasks. After perfecting the predictor with predicting acceptance of scientific articles and designing rewards for simulated robots, they used it to guide the improvement of new agents tasked with rating mathematical solutions. Of the 100 exercises reserved for final evaluation, these agents exactly matched expert ratings 55% of the time, compared to 51% for early predictors and 47% for hyperagents.
Better performance does not mean lower cost. We must consider how much it will cost to achieve these improvements. In programming tests, the new method required an estimated average of 170 million tokens to complete 40 rounds, while HyperAgents used about 49 million. We are talking about artificial intelligence models that process or generate text units. The difference is substantial, though it doesn’t tell the whole story. When the researchers matched the token budgets of both systems, the method with prediction continued to outperform on the cumulative performance measure with a 10.9% advantage across the entire process. Of course, the study does not provide a complete comparison of total computational costs.
Which is yet to be proven. It’s one thing to pass a programming test and another to entrust a system with changes to the software we use every day. These research tests, available as pre-publication, do not confirm that the method works the same in real projects or that it eliminates the need for human supervision. Nor do they demonstrate that these agents can replace professionals in charge of developing or evaluating programs. But they show that anticipating potential outcomes can help guide change. Thus we return to the initial dilemma from another perspective: to improve their own performance, these agents need more than the ability to modify instructions and code.
Featured Image | Jakub Zerodzicki
In Xataka | Getty is the first big company in the past to fall due to AI: it won’t last if others don’t learn
