Hierarchical Orchestration Router

Hierarchical Orchestration Router

Overview

Discovery and retrieval find the right tool. Orchestration coordinates them at scale. As tool counts grow from dozens to hundreds and agent counts from one to many, you need infrastructure that routes requests, manages failover, and enforces governance. This skill composes three ideas from the chapter's "Orchestration at Scale" section.

Inversion — one orchestrator, not thousands of tools. Instead of exposing every tool to the agent, expose exactly one: an intelligent orchestrator that handles all complexity. The agent asks for an outcome; the orchestrator decides how to achieve it. This is what lets traditional SaaS answer "why did we lose deals last quarter?" in seconds instead of through menus and reports.

Hierarchical routing. Organizations don't just have more tools — they have multiple MCP servers across departments (Sales, Finance, Operations), each managing hundreds of tools. The router classifies a query into a domain by semantics. High confidence (> 0.8) routes to that domain's orchestrator; low confidence means the query spans domains, so it invokes cross-domain orchestration (chapter Example 6-11). This buys fault isolation (a logistics outage does not stop sales), scalable governance (global vs domain-local policies), and progressive disclosure (users see only capabilities relevant to their context).

Functional clustering for resilience. Within a domain, tools with similar function are clustered for intelligent failover. Baidu's AI Search Paradigm embeds tools by what they DO (DRAFT-refined docs + usage patterns), then K-means++ groups them into functional toolkits. When the primary tool is overloaded, the orchestrator fails over to a functionally-equivalent alternative from the same cluster — a "Search Toolkit" of Baidu AI Search / ArXiv MCP / Perplexity / OpenAI WebSearch — adapting parameters as it fails over. No single point of failure.

When to Use

• Multiple departments/domains each own many tools (or MCP servers)
• Queries arrive that may belong to one domain or span several
• You need failover: when one tool is overloaded, route to an equivalent
• You are turning a legacy multi-tool surface into a single natural-language entry

Phrases that invoke this skill: "route this query", "which domain owns this", "cross-domain", "one orchestrator", "fail over to an equivalent tool", "functional clustering", "hierarchical orchestration".

When NOT to Use

• A single-domain system with a handful of tools — routing overhead exceeds benefit; use rag-mcp-tool-selection directly.
• As tool retrieval within a domain — this routes to a domain and names the functional clusters; picking the specific tool inside a domain is rag-mcp-tool-selection / the gateway.
• As a security boundary — per-domain access control (which agent may reach which domain) is a separate authentication layer; this is coordination.
• When there is no failover set — a domain with one tool per function has no functionally-equivalent alternative; the skill correctly reports a single point of failure rather than inventing one.

Process

Rationalizations

Red Flags

• Every query routes cross-domain. The domain keywords are too generic or overlap heavily; sharpen them so a clear match scores near 1.0.
• A cross-domain query routes to a single domain with high confidence. The confidence metric is not margin-based; a spanning query must depress confidence.
• Functionally-unrelated tools land in one cluster. The clustering key (shared function) is too coarse — tighten key_topics so groups do not bridge.
• Failover returns a tool from a different function. The cluster lookup is broken; the alternative must be a same-cluster peer.
• CLI --help exits non-zero. SKILL.md / CLI mismatch; multi-harness invariant broken.

Non-Negotiable Verification

1. Run the benchmark battery.

   python cli.py benchmark
   

   Confirm single-domain queries route to the correct domain at high confidence, the spanning query is detected as cross-domain (< 0.8), and failover of baidu_ai_search lands on a same-cluster search tool.

2. Prove functional clustering is clean.

   python cli.py cluster
   

   Confirm the Search / Metrics / Logs toolkits are distinct (metrics tools do not merge with logs tools).

3. Prove the single-point-of-failure report. Add a tool with a unique function to the config and run python cli.py failover <that-tool> — it must report no alternative, not invent a cross-function one.

4. JSON round-trips.

   python cli.py orchestrate "..." --json | python -c "import json,sys; json.load(sys.stdin)"
   

Security Posture

• Prompt injection. Domain keywords and tool topics are author-controlled; treat them as untrusted if sourced from external registries. Routing decisions are deterministic given the config — a malicious keyword set could misroute, so the config is a governed artifact.
• Data exfiltration. No network calls in lib.py. The # TODO(production): seams in identify_domain (embedding classifier) and cluster_tools (K-means++ over embeddings) mark where real ML backends attach.
• Privilege escalation. Cross-domain routing must still honor per-domain access control — routing a query to Finance does not grant the agent Finance tools; that gate is mcp-gateway-two-meta-tools / IAM. This skill decides WHERE a query goes, not WHETHER the agent may go there.

Composition

• Sits above rag-mcp-tool-selection / mcp-gateway-two-meta-tools: the router picks the domain; those pick the tool within it.
• Consumes draft-tool-trust-verifier output: functional clustering embeds tools "by what they do, not what they claim" (DRAFT-refined representations), and the reliability failover ordering is a performance-based trust score.
• Feeds the Enterprise AI OS control plane — the router is the semantic-search
  • action-governance step in the chapter's EnterpriseAIOS.handle_agent_request.
• Pairs with federated-context-governance — hierarchical routing needs a coherent per-domain context to route into.

Source Attribution

Distilled from Agentic GraphRAG (O'Reilly), Chapter 6 — Tool Orchestration, "Orchestration at Scale": "The Intelligent Orchestrator", "Hierarchical Orchestration: When Scale Demands Structure" (Example 6-11), and "Functional Clustering: Resilience Through Redundancy". Named references:

• MCP Gateway hierarchical orchestration — Enterprise Orchestrator → domain → tools; identify_domain confidence 0.8 routing threshold
• Baidu AI Search Paradigm — functional clustering via K-means++ over usage-pattern embeddings; Search Toolkit failover example