32 AI agents sped up NumPy 18.9x, and teamwork was uneven
In a test run by Replit, a team of 32 Claude agents made the NumPy scientific library 18.9 times faster in two hours. A single Claude agent managed 3.7x. A team of 32 OpenAI agents did worse than its own best solo agent. The result hints that adding agents does not help automatically: how well they coordinate may count for more than how many there are. The figures are as reported and come from one test, not an independent benchmark.