Moonshot AI Previews Mira, a Group-Chat-Native Agent Harness for Teams

Moonshot AI extends the agent harness from computers to teams

Moonshot AI published a new Kimi resource on August 27, 2026 that describes Mira, a group-chat-native agent harness intended to place an AI coworker inside team conversations. Mira is not presented as a broadly released product in the material: Moonshot is inviting users to apply for beta access.

The design starts from a different assumption than a conventional coding or desktop agent. Instead of treating one computer, terminal or browser as the primary environment, Mira treats each group chat or direct-message space as a separate operating context with its own memory and permissions.

Three design ideas stand out

Moonshot describes context isolation as a core primitive: each group or private chat is an independent space. The resource explains the intended boundary as keeping the agent's actions and speech inside the current space while allowing information from other spaces only when explicitly authorized.

The second idea is timing. A team agent cannot simply respond to every message or wait for a command in the same way as a personal chatbot. Mira is designed around deciding when it should participate in an ongoing group conversation, making proactive timing part of the harness rather than an afterthought.

The third idea is knowledge scope. Moonshot frames organizational memory as something that accumulates over time even though individual execution environments can be reset. That makes permission boundaries and authorized cross-space retrieval central to the product concept.

Why Moonshot is calling this a harness problem

The same resource argues that agent performance depends heavily on the software harness around the model: prompt assembly, tool exposure, execution, observation handling, stopping conditions, permissions and trace logging. Moonshot positions its existing Kimi stack around that idea, including browser control, approval-before-action boundaries and Agent Swarm orchestration.

Mira applies the same logic to Agent × Team rather than only Agent × Computer. The important claim is architectural, not benchmark-based: a team agent needs isolation, permissioning, memory and participation rules that are different from a single-user coding agent.

Preview status and what remains unknown

Moonshot's page includes an application for Mira beta access, so this should be treated as a preview/private-beta-stage system, not as generally available software. The public material does not establish broad availability, pricing, enterprise compliance coverage, supported regions or a final feature set.

That distinction matters because group-chat agents can touch sensitive workplace context. The most important questions for a production deployment will be how permissions are enforced, how data is retained, what audit controls exist, how cross-space retrieval is authorized and how administrators can limit proactive actions. Those details require product documentation beyond the current preview description.