Anthropic’s research site has released three MAMUBench environments containing 74 scenarios for agents working on behalf of different users. Shared budgets, clinic calendars and personal-assistant tasks let researchers examine competing requests. The code release follows a September preprint and makes the experimental settings available for closer inspection and reproduction.
Anthropic has made three MAMUBench evaluation environments available through its research site. This single research document describes a release containing 74 scenarios for agents acting for different users. The environments recreate shared spending limits, clinic appointment calendars and personal-assistant tasks. Publication of the code follows a September 30 preprint, allowing researchers to inspect and reproduce the simulated settings.[1]
The experimental design compares one coordinator receiving all users’ requests with separate agents representing each user. Those teams can run with or without a communication channel. Users may compete over the same budget, time slot or order, so finishing a group task and preserving an individual request are separate concerns. The released personal-assistant environment includes 25 scenarios involving orders or bookings.[1]
The setup distinguishes a message being sent from the deciding agent actually reading it before committing an action. A tested safeguard blocks commitment while unread messages remain. These are simulated tasks: the clinic environment involves no real patients. Some consent and fabrication labels were assigned by language models without validation against human judgments. The package provides experimental code and settings rather than new measurements of patient or customer outcomes.[1]
