Simular Agent S4 Evolves Interfaces, Beats Opus 4.8 on OSWorld

Simular's new paper has an agent writing its own tools. In Evolving Agent Interfaces, the team introduces Agent S4, which pairs Qwen3.8-27B with a repository of interfaces the system builds and revises itself; on OSWorld V2 the setup reaches 21.3%, which the abstract reports as "3.8 times the same model with fixed interfaces."

That 21.3% sits above Claude Opus 4.8 at 20.6% and GPT-5.5 at 13.0%, "at about 4.3% and 1.5% of their respective costs." On ScienceBoard, the same evolved-interface setup lifts the open model from 41.4% to 58.6%.

The mechanism is where the paper is most pointed. A meta-agent reads task-agent trajectories on synthetic training tasks "without success labels" and rewrites the interfaces and usage knowledge stored in a shared repository. Population-based evolution then "selects which repositories to develop further by their performance on validation tasks the meta-agent never sees, favoring changes that generalize beyond the seen tasks."

The authors frame the underlying frustration plainly: "An agent may type a spreadsheet cell by cell when a single line of code would do, or print all two thousand lines of a file through a shell to find one function." Hand-built interfaces, they argue, cannot keep pace with new applications and models, so interface design "should itself be an agent capability."

It lands in a dense week of open-weight agent releases on our tracker, alongside Sage-A1.5 and Specter the same day. The abstract reports results on two benchmarks only, and does not break out which specific interface changes drove most of the lift, nor the compute cost of running the meta-agent itself.