Anthropic MAMUBench: Reading pending messages recovers 73.1% of failed multi-agent orders

·2026.11.13 02:31

Key point

A guard blocking checkout until agents read pending messages recovered 73.1% of failed episodes for Opus 5, compared to 45.5% from simple reruns.

In Anthropic's MAMUBench simulated group orders, the critical request reached the final-ordering agent in 94–99% of committed runs but was present in its context at checkout in only 47–67%.
Adding a guard that blocked checkout until the agent read all pending messages recovered 73.1% of previously failed episodes for Opus 5 and 62.3% for Sonnet 5.
Rerunning the failed episodes without the guard recovered 45.5% for Opus 5 and 30.3% for Sonnet 5.
The results apply specifically to the rerun failed episodes in simulated tasks, not all orders, and were tested on Opus 5 and Sonnet 5 rather than Fable 5.2.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.