Benchmarking Computer Use
Canine
Benchmarking Computer Use
Running OSWorld on Canine's agent computers
What happened when OSWorld's tasks were run on a real, stateful Canine agent computer: how the benchmark works, the harness that emulates its VM, every way the harness got the score wrong, and what GPT-5.1, Claude and Qwen3.8 27B showed about the agent. Six chapters, with the scores and costs of 140-odd graded runs.