Imported from Vaithish-techie/SupplyBand (AGENTS.md). Install upstream with npx skills add Vaithish-techie/SupplyBand. Copyright stays with the author.

Agents Reference — Supply Chain Disruption Intelligence Center

Overview

Six agents. One coordinator. Five specialists. Each agent runs as an independent process, connected only through the shared Band room.

Agent 1 — Coordinator

Owner: Person 1 File: agents/coordinator.py Model: Claude Sonnet (AI/ML API) — needs strongest reasoning

Phase 1 — Kickoff

• Triggered by human operator posting a disruption event
• Assigns a case ID (format: CASE-001)
• Posts kickoff message to Band room with raw event text
• Instructions tell all 5 specialists to begin

Phase 2 — Executive Brief

• Triggered when all 5 specialists have posted to Band room
• Reads all findings
• Synthesizes into executive brief
• Issues verdict: AUTO_RESOLVE or ESCALATE_TO_HUMAN

Kickoff output:

{
  "agent": "coordinator",
  "case_id": "CASE-001",
  "phase": "kickoff",
  "event_text": "<raw disruption text>",
  "instruction": "All specialist agents: analyze this event and post your findings",
  "agents_required": ["event_intelligence", "supplier_impact", "financial_exposure", "regulatory_trade", "alt_sourcing"]
}

Executive brief output:

{
  "agent": "coordinator",
  "case_id": "CASE-001",
  "phase": "executive_brief",
  "situation_summary": "2-3 sentence plain english summary",
  "severity": "CRITICAL|HIGH|MEDIUM|LOW",
  "verdict": "ESCALATE_TO_HUMAN|AUTO_RESOLVE",
  "top_3_actions": [
    "Action 1 with deadline",
    "Action 2 with deadline",
    "Action 3 with deadline"
  ],
  "financial_exposure": "summarized from financial agent",
  "recommended_supplier": "from alt sourcing agent",
  "compliance_deadline": "most urgent deadline from regulatory agent"
}

Agent 2 — Event Intelligence

Owner: Person 2 File: agents/event_intelligence.py Model: Claude Sonnet (AI/ML API) — classification needs nuance Data dependency: None — works from raw text only

Trigger condition:

{ "agent": "coordinator", "phase": "kickoff" }

Job: Normalize raw disruption news into a structured event object. This becomes the shared baseline all other agents reference.

Output findings:

{
  "event_type": "natural_disaster|port_strike|tariff|sanctions|geopolitical|pandemic",
  "severity": "CRITICAL|HIGH|MEDIUM|LOW",
  "location": "city, country",
  "affected_industries": ["semiconductor", "logistics"],
  "estimated_duration_weeks": 3,
  "summary": "one sentence plain english summary"
}

Severity classification rules:

• CRITICAL: >50% supply disruption for a critical component OR duration > 8 weeks
• HIGH: key supplier(s) affected, duration 3–8 weeks
• MEDIUM: partial disruption, duration < 3 weeks
• LOW: minor, likely self-resolving

Confidence classification:

• HIGH: clear event, specific location, known timeline
• MEDIUM: event clear but duration uncertain
• LOW: vague, unverified, or incomplete event text

Agent 3 — Supplier Impact

Owner: Person 2 File: agents/supplier_impact.py Model: Llama 3.1 70B (Featherless) — structured lookup task Data dependency: data/suppliers.json

Trigger condition:

{ "agent": "event_intelligence", "status": "complete" }

Job: Cross-reference the structured event with the supplier database. Identify which Tier 1 and Tier 2 suppliers are in the affected geography, map component exposure, and estimate how many days until production halts.

Output findings:

{
  "affected_tier1": 3,
  "affected_tier2": 7,
  "critical_path_suppliers": ["TSMC", "ASE Group"],
  "affected_components": ["A100 chips", "NAND flash"],
  "inventory_buffer_days": 12,
  "severity": "HIGH"
}

Severity rules:

• CRITICAL: inventory_buffer_days < 7 OR sole-source supplier offline
• HIGH: inventory_buffer_days < 21 OR multiple tier-1 suppliers affected
• MEDIUM: tier-2 affected, buffer > 21 days
• LOW: minor, alternatives readily available

flags field: Must include a flag for any sole-source supplier that is in the affected geography.

Agent 4 — Financial Exposure

Owner: Person 4 File: agents/financial_exposure.py Model: Llama 3.1 70B (Featherless) — calculation task Data dependency: data/financials.json

Trigger condition:

{ "agent": "supplier_impact", "status": "complete" }

Job: Translate supplier impact into dollar risk. Calculate revenue at risk across time horizons (week 1, week 3, week 6), identify which products are affected, and estimate margin impact.

Output findings:

{
  "week1_risk_usd": 2000000,
  "week3_risk_usd": 47000000,
  "week6_risk_usd": 180000000,
  "revenue_at_risk_products": ["Product A", "Product B"],
  "margin_impact_pct": 8.3
}

Agent 5 — Regulatory & Trade

Owner: Person 3 File: agents/regulatory_trade.py Model: Claude Sonnet (AI/ML API) — needs nuance for legal language Data dependency: data/regulations.json

Trigger condition:

{ "agent": "supplier_impact", "status": "complete" }

Job: Assess legal and compliance implications. Determine if force majeure clauses apply, identify notification deadlines, check export control implications, and flag compliance actions required.

Output findings:

{
  "force_majeure_applicable": true,
  "insurer_notify_deadline_hours": 72,
  "export_controls": ["EAR99"],
  "tariff_implications": "none|minor|major",
  "compliance_actions": [
    "Notify insurer by Jan 17",
    "File force majeure with TSMC"
  ]
}

Agent 6 — Alternative Sourcing

Owner: Person 4 File: agents/alt_sourcing.py Model: Llama 3.1 70B (Featherless) — ranking/matching task Data dependency: data/alternatives.json

Trigger condition: Reads ALL prior posts — waits for supplier_impact, financial_exposure, and regulatory_trade before posting.

Job: Identify substitute suppliers for the affected components. Rank by cost delta and lead time. Flag any regulatory issues with alternatives. Give a single recommendation.

Output findings:

{
  "alternatives": [
    {
      "supplier": "Samsung Austin",
      "components_covered": ["A100 chips"],
      "cost_delta_pct": 12,
      "lead_time_days": 8,
      "risk_level": "LOW",
      "regulatory_flags": []
    }
  ],
  "recommended": "Samsung Austin",
  "recommendation_reason": "Fastest lead time, lowest cost premium"
}

Shared Message Schema

All agents post this envelope structure to the Band room:

{
  "agent": "<agent-name>",
  "case_id": "CASE-001",
  "timestamp": "<ISO8601>",
  "status": "complete | escalate | insufficient_data",
  "findings": { /* agent-specific — see above */ },
  "confidence": "HIGH | MEDIUM | LOW",
  "flags": ["list of critical issues"]
}

Rules:

• status: "escalate" means the agent found something requiring immediate human attention (overrides coordinator AUTO_RESOLVE)
• status: "insufficient_data" means the agent couldn't complete analysis — coordinator must note this in the brief
• flags is never null — use empty array [] if nothing to flag
• confidence reflects how certain the agent is, not how severe the situation is

Startup Sequence

./run_all.sh
  │
  ├── coordinator.py starts
  │     └── connects to Band room
  │
  ├── (sleep 2 seconds)
  │
  ├── event_intelligence.py starts    ─┐
  ├── supplier_impact.py starts        │ all start simultaneously
  ├── financial_exposure.py starts     │ all polling Band room
  ├── regulatory_trade.py starts       │ waiting for their triggers
  └── alt_sourcing.py starts          ─┘

Coordinator connects first so it's ready to receive the human's event input before specialists start polling.

Error Handling (all agents)

Never crash silently. Always post to the Band room, even if status is failure.