The model stayed put. The buttons changed.
Simular's Agent S4 treats the interface repository as the thing worth evolving. On OSWorld V2, that choice is enough for an open 27B to clear two frontier APIs on binary success while spending a few percent of their reported dollars per task.
Interfaces become the evolutionary object while model and harness stay fixed
Agent S4, described as an autonomous computer-use scientist, does not solve benchmark tasks itself. A task agent attempts synthetic training tasks through a candidate repository of interfaces and usage notes. A meta-agent reads those trajectories without success labels, then mutates or crosses repositories: adding, revising or removing interfaces and the knowledge of when to call them. Population selection prefers parents with higher held-out validation scores. The meta-agent never sees validation tasks, their trajectories or their scores, which is the paper's main generalisation guard.
An interface here is the request and observation language for some part of the computer: GUI clicks and screenshots, shell commands and text, spreadsheet bulk edits, ChimeraX residue colouring, and so on. The harness still owns invocation, budgets and permissions. Only the repository changes across the evolutionary loop. That framing is the novelty claim: interface design as an agent capability learned from experience, rather than a hand-written framework release every time a new failure mode appears.
OSWorld V2 binary success and a ScienceBoard lift on the same 27B
On the full 108-task OSWorld V2 set, Qwen3.8-27B with the evolved repository reaches 21.3% binary success and 46.9% partial in the no-scratch-VM condition the authors highlight for binary score. The same model with fixed interfaces sits at 5.56% binary. GPT-5.5 batched lands at 13.0% binary and Claude Opus 4.8 batched at 20.6%, with reported costs of about $25.50 and $72.40 per task against roughly $1.10 for the evolved 27B run. Those ratios are the paper's "about 4.3% and 1.5%" cost claims.
A scratch-VM condition, where the meta-agent can open apps and try code while building interfaces, reaches 19.44% binary and 51.06% partial, with more new interfaces (12 versus 4) and many more extensions. Evolution without the scratch VM still wins on binary success.
On ScienceBoard's 169 tasks across six scientific applications, a fresh evolution with the same task agent and seed repository raises success from 41.4% to 58.6%. The Simular post also reports 99 of 169 solves for the evolved repository against 93 for the seed under their counting.
Same open weights, fixed harness, and API numbers taken from batched runs
The cleanest comparison is Qwen3.8-27B fixed versus evolved: same model, same harness family, different repository. Frontier API rows are batched evaluations with their own interface stacks and price tags; they are useful cost and score context, yet they are cross-stack rather than ablations. MiniMax M3 and Kimi K2.6 fixed-interface rows sit far below on binary success despite larger parameter counts, which supports the paper's thesis that interface mismatch burns capacity. ScienceBoard compares an evolved repository against the human-designed Qwen CUA interface on the same model, which is fairer than mixing closed APIs. Training tasks are synthetic and checked for non-overlap with the benchmark; readers should still treat "never saw the benchmark" as process claim backed by the authors' pipeline description.
Production desktop fleets, adversarial apps and long-horizon cost under live users
The paper does not claim that evolved interfaces transfer unchanged into arbitrary enterprise golden images, multi-user VDI farms or apps that change weekly under IT policy. It does not publish a full threat model for interfaces that can run shell or browser requests with stored credentials. Dollar figures are per-task research accounting on OSWorld V2, so they do not settle your cloud bill once you add human review, VM pooling and failure retries. Whether the same evolutionary loop stays stable when the task distribution drifts after deployment remains open. Cross-benchmark reuse of a single evolved repository without a fresh ScienceBoard-style run is also outside the reported protocol.
Paper, Simular post and Agent-S repository are public
Primary artefacts are arXiv:2610.28601, the 21 October 2026 Simular article, and the open Agent-S code named in the paper. Evolved repositories and reproduction scripts should be read from that repository and the paper's release notes as they land; treat any missing weight or trajectory dump as unsettled until linked from those primaries. Licence terms follow the repository LICENSE file.
Trial interface evolution on a narrow app family before touching the model budget
Teams stuck improving computer-use agents by swapping GPT-class APIs should run a bounded Agent S4-style loop on one application family they already scrape trajectories for, keep the model pin fixed, and score only on a held-out validation slice the meta-agent cannot see. Prefer that experiment over another prompt-only pass if trajectory review already shows cell-by-cell GUI grind or screenshot-read shells. Buyers comparing OSWorld leaderboard rows should file the 21.3% open result next to cost columns, then ask vendors whether their demos ship fixed human interfaces or an evolving repository. Skip a full evolutionary stack if you lack synthetic-task generation and a validation split you are willing to keep blind.
Beside OO-Harness: tools the agent sees versus harness subclasses
Frontier's same-week OO-Harness coverage (arXiv:2610.28380) evolves specialised harness subclasses and control-loop structure. Agent S4 evolves the interface surface those loops call. Both reject a single frozen human design, and they compose rather than compete: a better harness still needs the right spreadsheet and shell primitives, and a rich repository still needs invocation and budget policy. Relative to earlier Agent-S generations and skill-distillation work, S4's contribution is making the repository itself the evolutionary genome under verifier-free synthetic experience.