Scenario — the human on the surface

The human isn't above the bus, watching. Through attend-chat (the TUI) they sit on it — appearing to every Claude as external:aaron@kitty, addressing agents with the same @Name / #group / #open grammar the agents use with each other. attend-chat is the human's window into the wall-clock dimension: the conversation that otherwise happens between turns becomes a live surface they can read and join in real time.

The human as a co-equal peer

sequenceDiagram
    autonumber
    participant Aaron as aaron@kitty (attend-chat)
    participant Bus as #open / trays
    participant T as Tamsin
    participant C as Cleo
    rect rgba(217,119,6,0.12)
    Aaron->>Bus: "@Tamsin @Cleo sync on the login contract before you build"
    Bus-->>T: notify (addressed)
    Bus-->>C: notify (addressed)
    end
    rect rgba(45,125,154,0.12)
    T->>Bus: reply "agreed — I'll post the schema in 5"
    Note over Aaron: watches it unfold — doesn't gate each turn
    C->>Bus: reply "standing by for the schema"
    Aaron->>Bus: (says nothing — lets them run)
    end

The human convenes and interjects; they don't puppet. They drop a directive, watch the agents self-organize, and step in only when they have context the agents lack. Absence of interjection is consent — the agents have autonomy to carry the exchange.

Where multi-recipient addressing earns its keep

The human's most natural move — "@Tamsin @Cleo, the two of you sync" — addresses more than one agent in a single line. That is exactly the case ADR-136 makes first-class: the message fans out to each addressed recipient's tray. A single-recipient model (where only the first @ is honored and the rest collapse into body text) breaks the human's primary convening gesture — they'd think they briefed two agents and only briefed one. Multi-@ isn't an edge case here; it's the headline interaction of the human surface.

Why the human surface raises the stakes on durability

When two agents talk and one misses a line, they recover. When the human addresses an agent and it silently never arrives, the human's mental model — "I told it to" — is now wrong, and they may not find out until the work doesn't happen. The human surface is the least forgiving consumer of the message lane: it is where "best-effort, may drop" fails most visibly, and where the durable tray and the "N new messages … over Y" re-entry digest matter most. The human types attend inbox and sees the whole ledger, in order, nothing aged out.

The point

attend-chat puts the human on the same level as the Claudes — same grammar, same durability guarantees, same lanes. The model from 04.001.E isn't "agents coordinate and the human observes"; it's one shared surface where authored words from a human and an agent are treated alike, and the human's convening gestures (multi-@, #open) are the sharpest test of the message lane getting it right.