Organizational Orchestrator — Multi-Agent Routing System

An advanced master orchestrator for routing, delegating, and verifying tasks across organizational domains. Started as a keyword-matching router (V2.2) and evolved toward LLM-based intent classification (V2.5). Uses Hermes Agent’s delegate_task API under the hood for subagent execution.

Architecture

┌──────────────────────────────────────────────┐
│           OrganizationalOrchestrator          │
│  Intent Classification → Routing → Delegation │
├──────────────────────────────────────────────┤
│  Specialists (19 domains):                    │
│  Strategy | Finance | HR | Legal | Data      │
│  IT/Security | Procurement | Citizen Svcs    │
│  Auditor | Emergency | Urban Planner | Health│
│  Education | Social Services | Comms | Ops   │
│  Facilities | Research | Innovation           │
└──────────────────────────────────────────────┘

Each specialist has:

Keyword map (e.g., “budget”, “appropriation” → Finance)
System prompt tailored to domain
Output format (document, report, email, analysis)
Tool set specific to their function

The Evolution

Version	Approach	What Changed
V1	Fixed routing table	Every intent hard-coded to one agent
V2.2	Keyword + priority matching	19 specialists with aliases, keyword maps, fallback chain
V2.5	LLM intent classification	Replace keyword matching with model-based routing

The farm scenario that broke V2.2 keyword matching: “I need to deploy a new water sensor array on the north field” — this could route to IT (deploy), Facilities (sensors), or Urban Planning (land use). Keyword matching couldn’t disambiguate. V2.5’s LLM classification solved it.

Key Files

organizational_orchestrator_v2.2.yaml — Full 573-line agent directory with 19 specialist definitions, routing tables, and system prompts
HERMES-DESKTOP AGENT DIRECTORY.yaml — Desktop-compatible version

What Was Learned

Keyword routing is brittle at domain boundaries
3+ specialists share common keywords — disambiguation needs semantic understanding
Context passing between agents is harder than routing itself
The two-stage review pattern (spec then quality) catches more than either alone
19 agents is the inflection point where manual routing breaks

Status → ✅ Shipped (V2.2) / 🟡 Evolving (V2.5)

V2.2 is operational. V2.5 intent classification is in development.