This is a great paper from AWS AI Labs. They show that multi-agent coding teams raised their test pass rate by 28.2% and cut wall time by 16.5% without changing the model. The authors achieve this by changing how the agents communicate. In free-form agent teams, shared