Simular lifted a 27B model past GPT-5.5 by changing its tools, not its weights. Fix the interface first.

Simular's Agent S4, released October 21, lifted a 27B open model past OpenAI's GPT-5.5 on desktop tasks by rewriting its tools, so fix your agent's interface before you buy a bigger model. In Simular's Agent S4 post, Qwen3.8-27B reaches 21.3% binary success on the 108-task OSWorld V2 benchmark. GPT-5.5 with batched actions scores 13.0%, and Simular puts the 27B setup at about 4.3% of its cost.

The model never changed. Agent S4 runs a second agent that reads the task agent's runs and rewrites the interfaces it works through. Simular's example is an agent typing a spreadsheet cell by cell when one line of code would do. With the evolved interfaces, the same 27B model scored 3.8 times higher than with fixed ones.

Simular's summary: Interfaces are an axis of improvement separate from the model and the harness. These are Simular's own numbers, not an independent replication. The direction still holds up for anyone running an agent today.

Here's what I'd do. Read a week of your agent's runs. Find the step it repeats over and over when one call would do, and give it that call.

A bigger model types the same cells faster. A better tool skips them.