@dataeng.ai: One agent per user lost to one coordinator. Thirty percent to sixty-four. Anthropic tested Opus 5 and Sonnet 5 in four environments: a shared API budget, a clinic calendar, personal-assistant group orders and bookings, plus a merge queue. The API budget, clinic, and personal-assistant environments (74 scenarios) are the MAMUBench benchmark; the merge queue's 3 scenarios are in the paper but not the benchmark. The coordinator won all four environments for both Claude models. In the personal-assistant tests, agents on the team overrode each other and made up claims about users. With AGENTS.md prompt rules, requests kept rose from 44% to 58% for Opus 5, while the coordinator hit 88%. A checkout gate helped too: blocking checkout until the committing agent read all of its pending messages recovered 73% of failed runs for Opus 5 (62% for Sonnet 5). In the API-budget test, the team reached 47% of the coordinator's performance for Opus 5. As teams grew from 4 to 16 agents, the share that ever acted fell from 66% to 10% for Sonnet 5. Sources: Anthropic Alignment Science, Nov 11, 2026. arXiv:2610.00583. MAMUBench at github.com/safety-research/MAMUBench. Full links in the blog post (link in bio). #dataengineering #ai #aiagents #dataengineer
dataeng.ai
Region: US
Saturday 14 November 2026 03:54:09 GMT
Music
Download
Comments
dataeng.ai :
Does your agent team have one coordinator, or one agent per user?
2026-11-14 03:54:26
0
To see more videos from user @dataeng.ai, please go to the Tikwm
homepage.