The evidence · a benchmark case study

A single Claude agent versus a Roscoe workflow. Only one kept its footing.

No tricks, no thumb on the scale. The same Claude model (Haiku) ran on both sides, and the single Claude agent was even allowed to write and run its own code to help. Here is what happened when the work got long, and when someone slipped in bad data.

The comparison is published in full: what each side was asked to do, what it produced, and where the single agent drifted once the task outlasted its attention.