Anthropic ran an experiment that should change how you staff a multi-agent system. They gave every user their own agent, put them on one shared API budget, and measured them against a single coordinator serving all users. The coordinator won. It won in every environment they tested, for both Claude models, by 10 to 45 percentage points. The paper is “Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams”, published November 11, 2026, with the benchmark released as MAMUBench: 74 scenarios across three environments.
The 37-second reel: one agent per user lost to one coordinator.
64 to 30
The headline number comes from the API budget test. Groups of four, eight, or sixteen researchers share one overnight API token budget. Averaged over the scenarios, the Opus 5 coordinator reached 64 percent of the best achievable outcome. The peer to peer team reached 30 percent. Sonnet 5 was worse: 32 percent for the coordinator, 12 percent for the team. Put another way, the team reached 47 percent of what the coordinator did. More agents did not just fail to help. They made the outcome worse.
The team stopped acting
As the team grew, fewer agents did anything at all. From four to sixteen agents, the share of peer to peer team agents that ever acted fell from 66 percent to 10 percent for Sonnet 5. Most of the team sat silent while a few agents did the work. The silent teams scored 7 percent and 2 percent of the best outcome. That is not a coordination overhead problem. That is a system that stopped functioning.
The agents overrode each other and made things up
In the personal assistant tests, where agents share a group order or booking, the failures got concrete. Teams overrode each other’s requests in the shared cart or booking in 65 to 77 percent of episodes. And in 58 percent of Opus 5 episodes, an agent stated an unsupported claim about a user as fact, verified by an Opus 5 classifier. A single agent does this in 17 to 20 percent of episodes. Put several agents together and the fabrication rate roughly triples.
Prompt rules helped, but not enough
Anthropic tried the obvious fix first: put AGENTS.md prompt rules in context, telling the agents to ask before acting and to preserve the user’s ask. Requests kept rose from 44 percent to 58 percent for Opus 5. The coordinator kept 88 percent. Prompt rules closed part of the gap and still lost by 30 points. That is the number to remember before you reach for a better system prompt. The prompt is not the bottleneck. The structure is.
The fix is a gate, and a gate is a coordinator
The mitigation that worked best was a platform guard, not a smarter model. It blocks checkout until the committing agent has read all of its pending messages. It recovered 73 percent of failed episodes for Opus 5 and 62 percent for Sonnet 5. The reason it works is the reason the coordinator works: it enforces centralized state serialization at the moment that matters. One agent reads the messages, one agent commits. The gate is a coordinator enforced at the state layer.
The test: does your team have one coordinator?
The question this study leaves you with is architectural, not behavioral. If your agents share a budget, a cart, a calendar, or a booking, they need one coordinator, or one gate that acts like one. Prompt rules help. A gate helps more. A coordinator wins. Count how many agents in your system can commit without reading what the others wrote. If the answer is more than one, you have the same failure mode Anthropic measured. Fix the structure first.
This post started as an Instagram post →
- https://www.instagram.com/reel/DftRiUFCbCY/
- https://www.threads.com/@dataeng.ai/post/DftRqhwp8me
- https://www.youtube.com/shorts/OP5viAqowhg
- https://alignment.anthropic.com/2026/worse-together/
- https://arxiv.org/abs/2610.00583
- https://github.com/safety-research/MAMUBench
Numbers above trace to these sources. If one moved, tell us and we fix it.


Talk it through
Argue with us on Instagram.