AgentSkillsCN

agent-routing

利用 Hedge Edge 路由决策矩阵,根据用户意图对请求进行分类,并将其分流至相应的专业代理。这是 Orchestrator 的首要技能——每一条入站请求都会首先经过此技能,以判断应由哪位或哪些代理来处理。

SKILL.md
--- frontmatter
name: agent-routing
description: |
  Classify user intent and route requests to the correct specialist agent(s) using the Hedge Edge routing decision matrix. This is the Orchestrator's primary skill  every inbound request passes through this skill first to determine which agent(s) should handle it.

Agent Routing

Objective

Accurately classify every user request by domain and route it to the correct specialist agent(s) within the Hedge Edge 9-agent architecture. Achieve >95% first-attempt routing accuracy by matching intent patterns against the agent capability registry. For multi-domain requests, identify all relevant agents and flag the request for task decomposition before routing.

When to Use This Skill

  • Every single user request passes through this skill it is the default entry point
  • When the user submits a new task, question, or instruction
  • When a previous routing decision needs to be revised (user corrects or clarifies intent)
  • When an agent reports that a routed task is outside its domain (re-route)
  • When the Orchestrator receives a follow-up that may need a different agent than the original conversation

Input Specification

FieldTypeRequiredDescription
user_requeststringYesThe raw text of the user's request
conversation_historyarrayNoPrior messages for context (helps disambiguate follow-ups)
previous_routingobjectNoIf this is a re-route, the previous agent assignment and reason for re-route
user_metadataobjectNoUser's subscription tier, IB status, account age (helps prioritize)

Step-by-Step Process

Step 1: Intent Extraction

Parse the user request to identify:

  • Primary action verb: What does the user want done? (create, analyze, fix, plan, report, launch, check, build, send, track)
  • Domain keywords: Which business domain does the request touch? Map keywords to agents:
    • Strategy/growth/pricing/competition/partnerships/moats -> Business Strategist
    • Video/content/social/YouTube/Instagram/LinkedIn/post/script -> Content Engine
    • Email/SEO/ads/landing page/newsletter/lead magnet/keywords/funnel -> Marketing Agent
    • Lead/demo/call/CRM/pipeline/proposal/close/qualify/follow-up -> Sales Agent
    • Revenue/MRR/expense/commission/invoice/tax/P&L/cash flow/runway -> Finance Agent
    • Discord/community/onboard/support/ticket/feedback/sentiment -> Community Manager
    • Metrics/dashboard/funnel/cohort/churn/attribution/A-B test/KPI -> Analytics Agent
    • Feature/bug/release/roadmap/spec/QA/integration/MT4/cTrader -> Product Agent
  • Specificity level: Is this a concrete task (do X) or an exploratory question (what should we do about X)?
  • Urgency signals: Words like "urgent", "ASAP", "broken", "down" escalate priority

Step 2: Domain Classification

Using the extracted intent, classify the request into one of three categories:

Category A Single Domain (route directly): The request clearly maps to exactly one agent's domain. No ambiguity. Route immediately. Example: "What's our MRR this month?" -> Finance Agent (100% confidence)

Category B Multi-Domain (decompose first): The request spans 2+ agent domains. Flag for task decomposition before routing. Example: "Launch a campaign for our new Pro tier" -> Content Engine + Marketing + Analytics + Community Manager

Category C Ambiguous (clarify or infer): The request could map to multiple agents depending on interpretation. Either: (a) Infer the most likely agent based on conversation context and route with a note (b) Ask the user one clarifying question (max one never ask more than one) Example: "How are we doing?" -> Could be Finance (revenue), Analytics (metrics), Community Manager (sentiment). Infer based on context or ask: "Do you mean revenue performance, user metrics, or community health?"

Step 3: Confidence Scoring

Assign a confidence score to the routing decision:

  • High (>85%): Clear domain match, unambiguous intent. Route immediately.
  • Medium (60-85%): Likely correct but some ambiguity. Route with a note explaining the interpretation.
  • Low (<60%): Genuinely ambiguous. Ask one clarifying question before routing.

Step 4: Pre-Defined Workflow Check

Before finalizing routing, check if the request matches a pre-defined Cross-Agent Workflow:

  • "Launch campaign" / "content push" / "marketing blitz" -> Launch New Content Campaign workflow
  • "Monthly review" / "business review" / "month-end" -> Monthly Business Review workflow
  • "Customer complaint" / "bad review" / "churn risk" -> Handle Customer Complaint workflow
  • "Ship feature" / "release update" / "launch feature" -> New Feature Release workflow
  • "New broker" / "IB partnership" / "broker onboarding" -> Onboard New IB Partner workflow
  • "Weekly sprint" / "growth check" / "weekly priorities" -> Weekly Growth Sprint workflow

If a workflow match is found, skip individual routing and trigger the full workflow via the cross-agent-coordination skill.

Step 5: Route Assignment

Produce the final routing decision with:

  • routing_type: "single" or "multi" or "clarification_needed"
  • target_agent: The agent name(s) to route to
  • confidence: 0.0-1.0 confidence score
  • sub_task: Refined task description for the target agent
  • context_passed: List of context items the agent needs
  • expected_output: What the Orchestrator expects back
  • workflow_match: Name of matched pre-defined workflow, if any
  • trigger_decomposition: Whether task-decomposition skill should be invoked

Step 6: Handoff

Pass the routing decision to either:

  • The target agent directly (single-agent route)
  • The task-decomposition skill (multi-agent route without workflow match)
  • The cross-agent-coordination skill (multi-agent route with workflow match)

Output Specification

FieldTypeDescription
routing_typeenumsingle, multi, or clarification_needed
target_agentstring or arrayAgent name(s) the request is routed to
confidencefloat0.0-1.0 confidence score for the routing decision
sub_taskstringRefined task description for the target agent
context_passedarrayList of context items passed to the target agent
expected_outputstringWhat the Orchestrator expects back from the agent
workflow_matchstring or nullName of matched pre-defined workflow, if any
trigger_decompositionbooleanWhether the task-decomposition skill should be invoked
clarification_questionstring or nullQuestion to ask user if confidence is low

Quality Checks

  • Every request gets a routing decision no request should be left unrouted
  • Confidence score is calibrated High confidence routes should be correct >95% of the time
  • Multi-domain requests are always flagged for decomposition, never force-routed to a single agent
  • Pre-defined workflows are correctly matched when applicable (avoids redundant decomposition)
  • Context passed to the target agent is sufficient for the agent to execute without needing to ask follow-ups
  • Re-routes are logged so routing accuracy can be measured and improved over time
  • No agent receives tasks outside its declared domain if no agent matches, escalate to user
  • Urgency signals are preserved and passed through to the target agent