Evolving Agent Interfaces

Evolving Agent Interfaces. By Saaket Agashe and seven others, from the University of California, Santa Barbara, Cornell University, and two other institutions. Published on the twentieth of October. Its code repository is worked out: about two hundred and fifty source files, with a readme and a licence.

Computer-use agents can only be as effective as the interfaces humans hand them. The authors give two pictures of what goes wrong. An agent may fill in a spreadsheet one cell at a time when a single line of code would do the job. Or it may print all two thousand lines of a file through a shell just to find one function. Today both the interfaces and the knowledge of when to use each one are built by hand, and hand-building cannot keep pace with new applications, new workflows and new models.

Their position is that designing interfaces should itself be something an agent can do. Agent S4 is an autonomous computer-use scientist that builds what the authors call Evolving Agent Interfaces, drawing on its experience of tasks.

At the centre is a meta-agent. It examines the trajectories of task agents on synthetic training tasks, without being told which attempts succeeded, and from them it constructs and revises interfaces, together with knowledge of how to use them, stored in a repository. That pairing answers both halves of the problem described earlier. The interfaces are the means an agent acts through, and the usage knowledge is the sense of when each one should be used, the very judgment missing from the agent that typed cell by cell. Neither is written by hand any longer; both are drawn out of what the task agents actually did.

Population-based evolution then decides which repositories to develop further. Many repositories are kept in play at once, and the selection among them is made by their performance on validation tasks that the meta-agent never sees. Judging repositories on unseen tasks favours changes that generalise beyond the tasks the meta-agent learned from, rather than changes that only fit them.

On OSWorld V2, Agent S4 running Qwen3.8-27B with the evolved repository of interfaces sets a new state of the art among open-source computer-use agents. It reaches a binary score of twenty-one point three percent, which is three point eight times the score of the same model with fixed interfaces.

That result is above GPT-5.5, at thirteen percent, and above Claude Opus 4.8, at twenty point six percent. It comes at about four point three percent of the first one's cost and about one point five percent of the second's.

On ScienceBoard, the evolved interfaces raise the success rate with the same model from forty-one point four percent to fifty-eight point six percent.

Here the interface stops being a fixed gift from humans and becomes something the system revises for itself. The authors place that alongside models and harnesses as a third lever: useful interfaces can be evolved from experience, and autonomous interface design becomes a source of improvement complementary to advances in the other two. The code is open-sourced.