Full presenter notes Show: Keep the FY27 map available. Open Competitor worker assignment , then the prepared cmux workspaces and the orchestrator skill’s routing sections.
Say: “One worker researches competitors. Another works through our historical information and product mappings. Each gets a bounded job and a file to return. The model needs those inputs and a common set of assumptions.”
The grid, 14:00 in this slot — build it on screen, 30 seconds: Start in one workspace with a single pane. ⌘D splits to the right. ⇧⌘D splits the new pane downwards. ⇧⌘D again on the left. You now have a 2×2 grid, built in three keystrokes in front of the room. Launch cx in one, muse in another and grok in the third; keep the fourth for yourself. Say what the room is looking at: three different companies' models, on one screen, about to answer the same question.
The brief — same question, three models, 2 minutes: Send all three the question they already watched you ask in section 6, with a rendering instruction on the end:
> “Compare Albus and Hermione on the upfront-payment treatment in the Lumos readout. Cite the source heading. Distinguish the question, the provisional view and any final conclusion. Write the answer as a single self-contained HTML card, no more than one screen, and open it in the cmux browser in this pane when you are done.”
The room knows the answer by now. They saw the heading in Meetings/2025-05-23 Lumos readout — DD findings.md: Albus asks a hypothetical, Hermione gives a provisional view and says the agreement and the related-asset treatment need checking, and nobody records a final conclusion. So the audience is the judge, and they can mark three answers without your help.
The output has to be visual. Three panes of scrolling terminal text is a reading exercise and it will lose the room. The rendering instruction is what makes this a demonstration: each pane resolves from a running terminal into a rendered card, and the grid ends as three finished answers side by side that can be compared at a glance. This is Execute from the COPE slide happening live — the phase renders the result, and it is HTML because both of us can read it. The agents open their own browser surfaces; the mechanism is cmux browser open 'file:///...' --focus true, which the installed cmux skill already documents, so no new tooling is needed.
Ask for a fixed shape so the three are comparable: Name the sections in the prompt — the question asked, the provisional view, the final conclusion, the source heading — so all three cards carry the same four rows. Divergence then shows up as a difference in one cell rather than three differently-worded essays. Consider asking for the source heading to be quoted verbatim, which makes a fabricated citation visible without opening anything.
What to point at when they land: Read the three cards side by side. Praise whichever cites the heading and reports that no final conclusion exists. Then look for the failure that matters: a card with something confident written in the final-conclusion row, for a meeting that reached none. That is the whole argument of this talk on one screen — the work came back fast, it rendered beautifully, and one of them was wrong in a way only someone who knows the deal would catch. Say that out loud. This is why the human sits at both ends.
The strongest thirty seconds in the session, if it lands: Three rendered cards, one of them quietly wrong, and a room that can see it themselves because you taught them the answer eight minutes earlier. Do not explain it. Let someone say it.
Why three vendors rather than three of the same: Ondrej's version of this demo runs three Claude Codes, which shows only that you can run three of something. Three different companies' models answering one deal question is the multi-model argument from part 1 and the A/B habit from the diff slide, proved live rather than asserted.
Preparation and recovery: Rehearse the whole sequence end to end, including the keystrokes and the render. Confirm each agent can write to a scratch directory and open a browser surface in its own pane; check that the card opens in the pane that produced it rather than stealing the focused one. Have all three rendered cards saved as HTML from a rehearsal run and ready to open directly; Ondrej's own live run produced a poor result in fifty seconds and the thing he actually presented was a prepared one. Verified 14 September: claude, codex, grok and muse are all installed and on PATH. Check each seat's model and remaining capacity on the morning. If a pane stalls for fifteen seconds, open that seat's rehearsed card from disk and keep talking; the comparison is the point and it survives a dead seat. Say that the card is a prepared one. Confirm Lumos is the open vault in every pane, because this question must never run against the real engagement.
Timing: This wants three to four minutes and the slot holds two. The cheapest trade is the AFK/remote cue below, which is thirty seconds of telling against three minutes of showing; the next cheapest is a minute from the graph section at 14:00. Decide at rehearsal and hold the 19:00 handoff boundary.
Say (the shape, 20 seconds): “I talk to one agent. I never talk to the others. That one lists the workspaces, reads what each worker has on screen, checks progress and hands out the next job, and it repeats that loop until the work is done. Four jobs run; my attention stays on one conversation.” Say this with the workspaces visible, then point at the one you are actually typing into.
Explain: The assignment names the inputs, owned output, runtime/model, acceptance check and stopping condition. Internal and external collection can run together. Their synthesis waits for the required inputs and reviewed classifications. A source gap goes back to the relevant worker.
Preparation: The embedded briefs are teaching examples; this build has launched no FY27 workers. Prepare two presentation workspaces with labelled rehearsal outputs and the actual runtime/model visible. Use the prepared assignment if those workspaces are unavailable. Keep source collection separate from writes to the business plan.
AFK / remote control · 30 seconds within this slot: Point to the assignment’s output and stopping condition. The agent can continue while you leave the keyboard. Show the already-paired remote view if rehearsed. In Claude Code, /remote-control shares the current terminal session with another device; the original computer must remain available. Check account eligibility and pairing before the talk. Remote Control documentation .
Show Competitor worker assignment here Show Orchestration: Chaos vs Control here