Why several agents and not one
One model that predicts, decides and acts fails in two different ways at once, so the capabilities are split into nodes.
In practiceA wrong forecast is absorbed by a buffer. A wrong action reaches a guest, a shift roster, or an aircraft.
The tempting shape is one large model that predicts, decides and acts. It looks elegant on a slide.
It collapses in practice, because forecasting and acting have different failure modes, different feedback loops and different audit requirements. A wrong forecast is absorbed by a buffer. A wrong action reaches a guest, a shift roster, or an aircraft. Splitting them is not a technical decision dressed as a product one; it is a product decision that happens to need a specific architecture.
So each capability became a node with a strict boundary, and the boundaries are enforced by the protocol rather than by discipline.
One scene, four signals
Guest experience is not an agent of its own, because the same twenty minutes belong to four domains at once.
In practiceCheck-in: occupancy, a transit strike outside, a returning guest's last request, and who is on the desk.
Nothing in this system performs check-in today. The scene is here because it is the shortest way to show why the boundaries are drawn where they are.
Check-in looks like a process, and optimising it as one gives you the metric a process offers: time to key. A fast, cold arrival is an excellent process and a bad welcome.
- TacetWhat is happening outside: a transit strike lands every arrival in the same twenty minutes.
- AetherixHow full the house is tonight, and what that does to the service that follows.
- AnimaWhether this guest has stayed before and what they asked for last time. For a first stay: nothing to recall, and it says so.
- PeritiaWho is on the desk, how loaded they are, how far into the shift.
This is why guest experience is not an agent of its own. An agent that owned it would have to own all four domains, and an agent that owns everything is not a bounded context.
Time to key can be predicted. Whether someone felt welcomed cannot. That gap is a property of the problem, not of the data, and it holds harder at check-out, which carries the payment and weighs most in memory.
The five agents
Four domain agents around one supervisory router, which holds no memory of its own.
In practiceSelect a mark on the figure to read what that agent owns, and what it is not allowed to own.
Four domain agents, one supervisory router at the centre. Orchestrator holds no memory of its own: state stays with the agent that owns it, and no agent talks to another. Select a mark to read what it owns and what it does not.
Aetherix executes. It does not orchestrate.
Five rules that hold it together
The boundaries are enforced by the protocol and by the server, not by anyone remembering to respect them.
In practiceThe property identifier is never a tool argument, so a model has no way to name another property, let alone read it.
One bounded context per node
A node that owns two domains starts arbitrating between them privately, and the arbitration stops being reviewable. Splitting is what keeps a decision attributable to a component. Concretely: Tacet can say a strike will push arrivals into one window, and Aetherix can decide to move two covers of prep forward, but neither can quietly do both and present the result as a single judgement.
MCP is the only doorway
Nodes do not import each other. Every exchange is a tool call, which means every exchange is loggable, auditable and priceable. It also means a node can be replaced without its neighbours knowing. Tacet's rules reach Aetherix as a tool response with a timestamp and a source, not as a shared table.
Until the orchestrator is built (Design), Tacet's output reaches Aetherix directly: the one documented exception, non-personal, read-only, logged. Target: every exchange goes through the orchestrator.
Tenancy is resolved server-side, never passed as an argument
The property identifier is established at every layer by the server and is never a tool input. What a model cannot name, it cannot cross. This is the difference between a boundary and a hope: a prompt asking for last week's covers "for the other property" has no argument to fill.
Memory is local first, federated second, and never raw
Each property learns from its own history. Cross-property learning moves only as anonymised aggregates. A thirty-room independent gets the benefit of a network, such as how a rainy Tuesday in the same district moves breakfast covers, without its own covers, guests or prices leaving the building.
ResearchFederation: Research, no substrate yet. Local memory is as described.
A fifth agent is an addition, not a redesign
The contract that governs exchanges between agents was designed to admit staff knowledge: what the people who work in a property know, which nothing captures today. That agent is named Peritia, the trade agent: a house's know-how, held by the house rather than by whoever is on shift. Not started, and no code.
What each node is allowed to learn
Autonomy is earned by feedback: a node that can be measured may adjust itself, one that cannot must be gated.
In practiceAetherix compares yesterday's forecast against the actual figure. The orchestrator has no outcome of its own to check.
The architecture is identical across nodes. What differs is whether reality answers back, and that decides how much autonomy each one earns.
| Node | Feedback available | Consequence |
|---|---|---|
| Aetherix | Manager replies, plus forecast compared against the actual figure | Can measure itself and gate on the result |
| Tacet | Rejections of its alerts, per coordinate | Learns local shielding without ground truth on the outcome |
| Anima | Guest signals, sparse and regulated | Learns slowly by design, and must forget on request |
| Orchestrator | None of its own | Judged only by whether the nodes it routed to were right |
A node with an outcome loop can be allowed to adjust itself. A node without one has to be gated before it ships, because nothing downstream will catch it.
Where the sector's twins stop
Hospitality digital twins model the building. None found models an operational decision before it is taken.
In practicePredictive maintenance and energy optimisation are mature. A staffing change simulated in advance is not.
Digital twins are mature and named elsewhere. Banks run market scenarios on them, Google models its data centres, Amazon its warehouses, and aerospace and automotive have used them for years to retire risk before touching hardware.
In hospitality they exist too, and they model the building: predictive maintenance, refurbishment simulated before the work starts, energy optimisation. Useful, and about walls.
Searching for one that simulates the impact of an operational decision before it is taken, a staffing change, an F&B order, a service adjustment, returns nothing. The sector models its assets and not its arbitrations.
This architecture is not marketed as a twin and does not simulate scenarios today. But cohort matching and the learning engine both point at that gap, and they arrived there from the operational side rather than from the vocabulary. Nothing was found, which is a finding, not a trophy.
Anima included: a guest memory, not a twin.
Verification and pointers
The statuses above are the claim. Here is where to check them.
In practiceThe meta-repository holds the contracts; the field note explains how a node earns the right to ship.
Public meta-repository: github.com/IvandeMurard/Hospitality-Multi-agent-Architecture, opens in a new tab.
The field note What is an LLM evaluation harness? Graders, protocols and closed loops for AI agents explains how a system earns the right to ship. This page explains what the systems are. Read that one next if the statuses raised the question of how they are checked.
Rose reads this note in full. She will not claim a node is further along than it is.