Why AI Coding Teams Stumble Even When Each AI Codes Well

AI is increasingly being used not as one helper but as a team of helpers. You hand a big coding job to several AI 'agents' — programs that can take actions on their own — and each handles a piece: one writes the database code, another builds the screen, a third tests it. But until now, we mostly graded these teams on one thing: did the final job get done? That's like grading a group project only on the poster, with no idea who actually did the work.

A research team has built a new measuring stick called AsynCodeBench. It takes 19 coding tasks from real software projects and maps out 52 'dependencies' — moments where one AI's work cannot move forward until another AI finishes something first. Then it checks two things: how many of those handoffs eventually worked, and when they happened during the run.

The finding is uncomfortable. Getting better at coding did not make AI better at collaborating. Across different AI models and sizes, the researchers found that improved solo coding performance often came with no improvement in teamwork — and the overall score could look fine even when the behind-the-scenes handoffs were quietly failing.

They also spotted a curious pattern they call a 'hopping window': successful coordination rarely builds up slowly. Instead, a burst of handoffs all resolve in a short stretch of the run — then things go quiet again. If you ever hand a big job to a team of AI helpers, this is the part to watch: the work only moves when the handoffs move.