Opus: Massive Online Agent Arena

Abstract

When humans and agents collaborate on complex shared tasks, one participant’s changes can invalidate another’s ongoing work or require others to review and incorporate new information before proceeding. Effective multi-actor coordination must mitigate the risk of actions based on outdated inputs changing shared state and keep required reviews and updates within declared limits, while preserving each participant’s ability to plan independently. We introduce Massive Online Agent Arena, a framework that uses collaboration interfaces to determine when proposed actions may change shared state. An arena comprises humans and agents, referred to as actors, together with a shared environment and interfaces that define what actors can observe and request, which actions require review, and which changes must be communicated to others. Actors can consult intentions, which declare planned actions, to anticipate interference and revise their plans.

Before applying an action, the environment verifies that its inputs remain current, its author has permission, and any required approvals are valid. It also reserves capacity for resulting updates that other actors must receive or incorporate. Review requests and updates count against declared limits on unfinished tasks, so completing them releases capacity for further actions. As participation and access change, the environment updates the requirements for applying actions and records which conditions leave proposals waiting.

This paper makes three contributions:

1. We introduce an arena framework designed to support agent to agent and human to agent collaboration at scale, with observable dependencies and enforceable interaction rules.
2. We develop intention sharing, action admission, and completion-based flow control, with conditions for valid approvals and bounded unfinished task weight under changing participation.
3. We demonstrate that the same coordination mechanism supports canvas and document operations while adapting rate restrictions to changing dependencies. Intention-guided selection reduces the average number of preparation attempts by up to 53.6% compared with a baseline that prepares actions without consulting other participants’ plans, while its effect on application latency varies across settings. We extend the evaluation to populations of more than 10⁶ simulated agents to examine how registry provisioning sustains progress at scale.