The agenda for our KI Back-2-Work Camp went up on our own page this week, and it settles a small question earlier notes left open: this is not a morning. It is two full days — 23 and 24 September 2026, start 09:30, closing on a Demo Day that opens at 16:30 — at Niederdorfstrasse 77 in Zürich, 420 metres and five minutes from Zürich HB. Round 05. Nine blocks, D01 through D09. Six to eight people; four seats left.

Read the agenda and one block stops you: D06, «GRAPH — vom Loop Engineer zum Graph Engineer» — from loop engineer to graph engineer. It is not a slogan. It names a real shift in how computational work is shaped. For years the unit of work was a loop: a script you run, wait on, and run again. The graph makes the unit a set of things that call each other — outputs feeding inputs, a dependency you walk rather than a sequence you repeat. Architects already live inside one such graph. Grasshopper, David Rutten’s dataflow canvas at McNeel, has never been a loop; it is a directed graph of components, upstream feeding downstream. D06’s claim is simply to carry that habit up from the geometry to the whole office.

D01 sets the frame one level higher — «BOOT — AEC-OS: das Betriebssystem deines Büros», the operating system of your office. The deliverable of the two days is not a script. Sprint 01, on 23 September, ends at D04, «Deine Office-Ops-App» — a thing you built, not one you watched. Sprint 02, on 24 September, spends a whole further day making it talk to model data (D05, «CONNECT — Modelldaten als API») and rolling it out (D08, «SCALE»). That second day is the one a morning could never buy.

Here is the honest part, and it is about the market, not the camp: most offices already own the tools. What they have is a browser tab of half-adopted subscriptions and no layer that ties them together. A demo adds one more tab. D05’s «Modelldaten als API» is the same instinct behind PAZGPT, our own Archicad-by-chat-and-voice tool: the model is data you can call, not a drawing you redraw. If you want to see where that thinking already ships, our Archicad add-ons are the shelf it sits on. D02, «PIPELINE — deterministische Bildpipelines statt Zufall», states the house position on generated images out loud — determinism instead of luck.

←TODAY: Two days, nine blocks, four seats — Round 05, 23–24 September 2026, 09:30 to 16:30. →3012: The office that survives runs on an operating layer, not a folder of loose scripts. Fulcrum: A graph you can walk beats a loop you rerun — the same reason Grasshopper won the geometry and will win the Büro.

Atelier: For a team living with AI this month, the move is not another trial. Pick the one workflow you repeat most — the render pass, the quantity take-off, the report — and draw it once as a graph of steps that depend on each other. Bring that sketch and one small workflow problem of your own; that is exactly what the two days are built to turn into a running app.

Hack: Map your office tasks as a graph, not a to-do list, and let the machine walk the dependencies for you. Say your report needs a render and a quantity take-off, and both need a cleaned model — that is three arrows, not a checklist.

graph = {"report": ["render", "quantities"], "render": ["clean_model"], "quantities": ["clean_model"]}
def resolve(node):
    for dep in graph.get(node, []): resolve(dep)
    print("run:", node)
resolve("report")

Run it and the office does clean_model first, then render and quantities, then report — the order falls out of the graph, not out of your memory. That is the loop-to-graph move in five lines; D06 is the same idea grown to the size of a Büro.

The agenda is public and the count is real. If the shape fits your desk, take one of the four seats and walk in with the workflow you most want to stop doing by hand.