Found first: a primary source the press has not covered yet.
A new benchmark tests whether coding agents can hold focus on a primary task after being interrupted by a second, unfinished task in the same repository. Across 7 agents and 19 Python repositories, Switch-SWE finds the average focus-task resolution gap is 11.5 percentage points, ranging from 6.0 to 15.0 pp depending on the agent.
What the source says
Shuochen Chang and colleagues built 400 matched session pairs: 200 uninterrupted baseline runs and 200 switched runs where the agent handled a second task before returning to the original, across 200 focus tasks from 19 Python repositories and 7 coding agents. Two interruption types are defined: "near" (different files, same parent directory) and "far" (different parent directories). Root causes include cross-task requirement confusion and loss of partial task state. A plug-in context-management module requiring no retraining recovers 5.5 to 15.5 percentage points of the gap by selecting task-relevant history and generating state summaries; the code is at github.com/Miraclecsc/Switch-SWE.
Why it matters
Production coding agents routinely handle interleaved tasks, but there was no benchmark measuring the resolution-rate cost of doing so. The 6.0 to 15.0 pp range across all 7 tested agents shows the effect is consistent. The plug-in module result means the problem is addressable without fine-tuning. That shifts it from a research finding to an engineering task for teams already running agents on real ticket queues.