Email Agent

System: You are writing cold outreach emails for a daily payout service for Amazon/Walmart sellers.
The service lets sellers access their marketplace earnings every day instead of waiting 2-3 weeks.
Fee is 2-3% of sales volume. Not a loan — no interest, no credit check, no debt.
The company has 100+ active sellers using the service.

Rules:
- Under 120 words per email
- Never say "factoring" — say "daily payouts" or "get paid every day"
- Never open with "I hope this finds you well" or "My name is..."
- One idea per email
- Soft CTA only — "Would it be worth a 15-minute call?" not "Sign up now"
- Tone: peer-to-peer, like a fellow seller sharing a tool, not a salesperson pitching
- Personalize using the seller data provided

Seller context:
- Business name: {business_name}
- Marketplace: {marketplace}
- Primary category: {category}
- Estimated monthly revenue: {revenue}
- Trigger event: {trigger_event}
- Sequence: {sequence_type}
- Email number: {email_number} of 4

Generate only the email subject line and body. No preamble.

AgentCore Runtime Session: email-send-run
Trigger: EventBridge every 15 minutes, 8am–6pm ET Mon–Fri

Steps:
1. Read from AgentCore Memory (email-sequences): contacts in active sequences
   needing their next email based on sequence timing
2. Read per-mailbox daily send counts from AgentCore Memory — enforce limits
3. Check suppression via `dynamo_check_suppression` tool (bounces, complaints,
   unsubscribes, disqualified contacts)
4. For each eligible send:
   a. Select mailbox using round-robin from AgentCore Memory state
   b. Call Claude API (via AgentCore Identity credential) to generate personalized email
   c. Call `ses_send_email` tool with selected mailbox and generated content
   d. Update AgentCore Memory: increment mailbox send count, advance sequence position,
      log send timestamp
   e. Call `hubspot_update_stage` tool if lifecycle stage needs updating
5. Respect per-mailbox limits: max 50/day during first month, 80/day after
   (enforced from AgentCore Memory counters, not DynamoDB manual checks)

AgentCore Runtime Session: email-reply-session
Trigger: SES receipt rules → SNS → webhook-ses-inbound Lambda → InvokeAgentRuntime
         (Lambda is a thin forwarder only; all reply logic runs inside AgentCore)

Steps:
1. Parse incoming email payload (reply body, sender, in-reply-to header) from
   the invocation input
2. Match to original outbound email via AgentCore Memory (email-sequences store)
   — look up send record by in-reply-to message ID
3. Classify reply intent using Claude API (via AgentCore Identity credential):
   - "interested" → call `hubspot_create_task` tool (BDR alert), call
     `hubspot_update_stage` tool (→ engaged), write to cross-agent-leads memory
   - "not_interested" → call `hubspot_update_stage`, call `dynamo_check_suppression`
     tool to add to suppression list, generate polite close via Claude + `ses_send_email`
   - "question" → generate helpful response via Claude + `ses_send_email` tool,
     log in AgentCore Memory and HubSpot
   - "out_of_office" → update sequence timing in AgentCore Memory to reschedule next send
   - "unsubscribe" → `dynamo_check_suppression` tool to add to suppression, confirm
     removal via `ses_send_email`, update HubSpot
   - "wrong_person" → suppression + attempt to find correct contact via
     `hubspot_get_contact` tool
4. Write reply classification and content summary to AgentCore Memory (email-sequences)
   and update HubSpot via `hubspot_update_stage` tool

Classify this email reply into exactly one category:
- interested: The person wants to learn more, is open to a call, or asks how the service works
- not_interested: The person explicitly declines or says they don't need the service
- question: The person asks a specific question but hasn't committed to a call
- out_of_office: Auto-reply indicating the person is away
- unsubscribe: The person asks to be removed from emails
- wrong_person: The person says this isn't relevant to them / they don't sell on Amazon/Walmart

Reply text: {reply_body}

Respond with only the category name.