A meta-agent reviews task trajectories and evolves a repository of interfaces and usage knowledge through population-based mutation and crossover.
In Simular’s OSWorld V2 evaluation, Qwen3.8-27B with the evolved repository reached 21.3% binary success across 108 tasks. The company reports 5.56% for the same model with fixed Qwen CUA interfaces, and a reported cost of $1.10 per task. Its comparison table lists GPT-5.5 at 13.0% and $25.50 per task, and Claude Opus 4.8 at 20.6% and $72.40 per task.
The repository can add application-specific or cross-application helpers, such as bulk spreadsheet edits and requests from a logged-in browser tab. Simular also reports ScienceBoard success rising from 41.4% to 58.6% with the approach. The source does not establish independent replication, performance outside these benchmarks, or that the reported cost comparisons hold in production deployments.