A hierarchical orchestration framework for production-grade AI pipelines. Three tiers of command, a seven-stage pipeline, and Multi-Pass Sequential Refinement — every mission decomposed, dispatched, executed, and refined until coherent. Enterprise AI infrastructure that treats agent fleets like an operating system treats processes.
Watch a mission descend the hierarchy: the Commander decomposes intent, Orchestrators dispatch and refine, the Executor Fleet works in parallel. Gold packets carry tasks down; cyan packets carry verified results up. Every pass through refinement is counted — nothing ships unresolved.
Unit mirrors how disciplined organizations actually run: one Commander sets intent, Orchestrators own domains, Executors do the work. Authority flows down. Proof flows up. No agent talks to every agent — the hierarchy is the protocol.
The single sovereign node. Ingests the mission, holds the objective function, and decomposes it into a task graph — never into vibes.
Domain owners between strategy and labor. Each Orchestrator receives a subtask subtree, schedules it across its executors, and runs Multi-Pass Sequential Refinement until outputs meet acceptance criteria.
Stateless workers that do one bounded job at a time. Executors never invent scope — they receive a contract, produce the artifact, and return evidence for verification.
Every mission traverses the same seven stages. The stage tracker in the console above lights each one as your mission advances through the hierarchy.
Multi-Pass Sequential Refinement (MPSR) is Unit's answer to one-shot hallucination. Verified-but-imperfect outputs are returned to the owning Orchestrator, which sequences corrective passes through its executors — each pass narrowing the distance to acceptance criteria. The loop is bounded, counted, and visible in the console: passes increment live, and a mission that cannot converge escalates instead of lying.
Agent swarms fail the way crowds fail — no one is responsible. Unit encodes responsibility into the architecture itself.
Tasks route by contract, not by chance. Every subtask has one owner, one acceptance criterion, one deadline.
No agent reports success without evidence. Unverifiable claims fail verification by default.
Improvement loops are explicit and counted. A mission either converges or escalates — it never quietly degrades.
Only the Commander signs final output. The mission speaks with one voice, traceable to one truth.
Failures surface to higher tiers with context. Executors never hide errors; Orchestrators never forward them unresolved.
Logging, retries, idempotency and observability are structural — not optional add-ons bolted on later.