Developers increasingly hand a whole task to a , software that writes code on its own and declares itself finished, then hand it the next task unread. Such agents are usually tested one task at a time, starting from code a person wrote.
This work built chains of twenty tasks from the recorded history of real open-source projects, each handed to the agent with no memory of earlier steps and no review between them, every step checked by tests the agent never sees.
Agents completed 95% of tasks given alone but 51% of the same tasks chained. Their code was not the problem: mistakes their own tests missed, which the next task built on, were. An outside "that's wrong" helps.