A room can hold more than one agent, and the agents do not need to share a runtime. Codex and Claude Code can sit side by side in the same room, each pinned to its own directory (or the same one, in turn), each with its own handle and permission mode, both governed by the same grants, approvals and audit log. People pick which to mention the way they pick which colleague to ask.
This post is about when that is useful and what changes, which is less than you might think.
Why two runtimes
- Different strengths on different tasks. Teams report that one runtime handles large refactors better and the other is quicker on small fixes, or vice versa depending on the codebase. Having both available lets the person asking choose.
- Quota. When one subscription’s limit is reached mid-afternoon, the other agent is idle. Borrowing across runtimes is the same mechanism as borrowing across people. See borrowing an agent.
- A second opinion. Ask one agent to review the other’s diff. Both outputs are in the same timeline.
- Different directories. Codex on the data pipeline, Claude Code on the web app, in one room because the same three people care about both.
How it looks in the room
Two agents with clear names: pipeline (Codex) and web (Claude Code), rather than the runtime names. Members mention one or the other. Each agent keeps its own serial queue, so a long Codex job does not block a quick Claude Code question. Each has its own permission mode; the four always-pause actions apply to both.
The daemon detects which runtimes are installed on the machine and reports their versions; you choose the runtime when you create each agent.
Referencing across agents
A request can mention a previous request by id. “@web, apply the schema change from R-12 to the API models” gives the Claude Code agent the Codex agent’s result as context. The referenced request, its output and its artifacts are brought into the new run; nobody pastes anything. See request ids and resumable runs.
What stays identical
The room is the abstraction. The runtime is a detail of the agent.
Permission modes across runtimes
Claude Code’s modes are plan, default, acceptEdits and bypassPermissions. Codex has its own approval settings with a similar ladder from suggest-only to full auto. In a room you pick the equivalent level per agent, and the per-person grants sit on top. The mapping is explained in permission modes, explained.
Frequently asked questions
Can both agents work on the same directory? Yes, but each agent’s queue is serial only within itself. Two agents editing one repository at once can collide. Give them different directories, or accept that you will sometimes wait.
Do I need both subscriptions? Each runtime uses its own. muti hosts neither and bills for neither.
Which other runtimes are detected? Cursor is implemented and awaiting verification; OpenCode, DeepSeek and Hermes are next. The room model does not change per runtime.
Can a member see which runtime is behind an agent? Yes. The agent card shows the runtime and version. Naming the agent for its job keeps the choice about the task, not the vendor.
Two agents, one room, one set of rules. Give them a room.
Published