Friday author's test: handing off coding tasks between agents cuts tokens 21%

On November 2, 2026, the author of Friday, an open-source app for running coding agents side by side, published a controlled experiment on DEV Community measuring whether handing a coding task from one agent to another actually beats starting from scratch. Friday supports this hand-off workflow, and the author tested it directly, for example when one agent hits its usage limit. The experiment covered 14 tasks and compared two signals: tokens consumed, and how much the receiving agent explored before making its first edit. Across the 14 tasks, a hand-off used a median of 21% fewer tokens than starting over, and the hand-off came out ahead in 13 of the 14. On exploration, the next agent explored half as much before its first edit: a median of 2 steps versus 4 for a cold start, with the hand-off lower in all 14 of 14 tasks. The author also reported that all 14 hand-off fixes passed the tests (the published excerpt cuts off there). By framing the work as a direct measurement rather than an assumption, the post argues that task hand-off can cut token use and reduce duplicated exploration, while leaving the magnitude and passing rate as the numbers readers can check against.