Anthropic reports Claude Opus 4 multi-agent 90.2% gain

Analyzing production data from multiple engineering teams, the author argues that multi-agent AI architectures should be used exclusively for read-only tasks like research and parallel reviews, where subtasks do not write shared state. The data shows that chaining writing agents together leads to geometrically decaying completion rates, significantly higher token costs, and compounding handoff failures compared to single-agent systems. Open article →