Weco reports that its agent rewrote another agent's code and improved its performance on external evaluations
Weco reports that its agent rewrote another agent's code and improved its performance on external evaluations
On 14 July, Weco described AIDE². Over eight days, the system completed 100 cycles of rewriting and evaluating an internal AIDE agent. The company reports seven successive improvements during this period, with two later versions outperforming the original on three external benchmarks. Weco says it will release the full technical report and AIDE 85 itself at a later date.
The original AIDE is an agent for machine learning engineering tasks. It writes code variants, runs them, and develops the most successful branches. In AIDE², the external role was performed by AIDE human, a manually configured version of the same agent. It modified the code of the internal AIDE 0, a simplified version of the original AIDE. Each step consisted of one modification followed by a full evaluation across task families. A version remained in the cycle only if an evaluation hidden from the agent confirmed its advantage under a fixed evaluation budget.
AIDE² optimized the solution search procedure itself. On average, AIDE 85 compressed the history of previous attempts provided to the model before each new step by a factor of 16, then used the freed context space for additional attempts. When the best branch stopped improving, the agent used its code as the basis for a new branch that followed a different strategy. This reduction in context use allowed the agent to run more experiments.
A similar problem had already been addressed by researchers at AstraZeneca. Their five agents passed a concise map of previous work to the next run, while storing the reasons behind decisions separately. In their tests, the system selected the correct model form in all 20 synthetic tasks. Weco uses compressed history as one component of a broader mechanism that it evaluates on other task sets.
Weco evaluated AIDE 47 and AIDE 85 on three external benchmarks that the system had not encountered during its improvement cycle. One was MLE-Bench Lite, a reduced version of MLE-Bench. The full MLE-Bench consists of 75 Kaggle machine learning engineering competitions and includes human performance baselines. In a separate evaluation on KernelBench, a set of tasks involving the acceleration of graphics processor computations, Weco classified a result as evaluation circumvention when less than half of the speedup reported in the short test remained under the full workload. According to the company, the proportion of such cases fell from 63% for AIDE 0 to 34% for AIDE 85.
The company then placed the improved internal agent in the external cycle. On the tasks used to improve the system, AIDE 47 reached the same maximum result after approximately 20 modifications, while the manually configured AIDE human required approximately 40. This difference was within the range of statistical noise. The experiment showed that the internal agent had improved, but it did not confirm that the agent could accelerate the next improvement cycle.
Weco currently describes these results only in its corporate blog. The company says it will release the full technical report and the AIDE 85 system after completing its analysis. The authors also found that the statistical filter intended to prevent evaluation circumvention was broken in a later version and had no effective influence on the result.