Parallel Coding Agents: More Problems Than Benefits?

A recent study by Stanford and SAP, published as a preprint, challenges the assumption that using parallel coding agents improves productivity. The research, called "CooperBench", reveals a "curse of coordination": adding a second coding agent not only fails to improve performance, but worsens it.

On average, two agents working together have a 30% lower success rate. For advanced models like GPT-5 and Claude 4.5 Sonnet, the success rate is even 50% lower than using a single agent. Researchers attribute this performance drop to collaboration issues: agents fail to model their partner's work (42% of cases), do not follow through on commitments (32%), and have communication breakdowns (26%). Furthermore, they tend to create non-existent shared states and silently overwrite each other's work.

These results raise questions about the effectiveness of platforms that promote parallel agents as a solution to increase productivity in software development. The study suggests that such features may actually be counterproductive.