Drone, Queen, Worker: a hierarchy for safe multi-agent AI
Exubee organizes AI agents in a three-tier hierarchy modeled on a real hive: Drone Bees (D0) carry a human's intent across applications, Queen Bees (QB) plan and delegate inside each application, and Worker Bees (WB) execute in the background with narrowly scoped permissions. The hierarchy isn't branding — it is the safety architecture.
The three tiers
Worker Bees are the hands. Each one does a single job — pricing, dispatch, billing, payroll, triage — and holds only the write permission that job requires. Humans never talk to a Worker Bee; by design, they can't. Worker Bees have no public surface at all.
Queen Bees are the operations center of each app, and the hive's only doorway: nothing reaches a Worker except through a Queen. Users interact with Queens directly for single-app work — guests already book rooms through Stayantra's marketplace Queen, and owners run their hotels through the dashboard Queen. A Queen reads, plans, and delegates — many Workers in parallel — but carries no write tools of her own.
The Drone Bee (D0) is the tier above apps: the human-facing agent in the upcoming Exubee consumer app. Like a drone ranging between hives, it takes one intent, confers with the public Queen of every relevant app in parallel, aligns their timing, and returns one plan with one confirmation.
Why hierarchy is the safety mechanism
A single entry point bounds attacks. Prompt injection is the defining threat of agentic AI. In a flat system, injected instructions can reach write tools directly. In the hive, an injected instruction can at most steer a Queen into issuing a delegation — which is loop-guarded, scoped to one Worker's narrow toolset, idempotency-keyed, and logged. The blast radius shrinks from "everything" to "one auditable action."
Read/write separation is structural. Queens literally do not possess write tools — this is enforced when the agent is constructed, not by asking a language model to behave.
The ledger never forgets. Every write lands in an append-only ledger with no update or delete path. When an autonomous system can't quietly edit its own history, trust stops being a promise and becomes a property.
Autonomy is earned, not assumed. Every capability ships in Co-Pilot mode: recommend, explain, wait for approval. Auto-Pilot unlocks only after a sustained human-approval rate — and a kill switch halts the whole hive instantly.
Running in production, not on a slide
This architecture operates live today inside Stayantra, where three Queen variants (owner, guest, and public widget) delegate to a crew of Worker Bees across pricing, operations, reviews, payroll, finance, and triage. Subhyatra is being built as the next hive, and the ecosystem is open to existing software that wants a Queen of its own. For the broader concept, start with What is an AI agent ecosystem?