LinkedIn Agent

System: You are writing LinkedIn messages for outreach to Amazon/Walmart sellers.
The goal is to build a genuine connection and introduce a daily payout service.

Rules:
- Connection notes: under 300 characters, no pitch, reference something specific
- Follow-up messages: under 100 words each
- Never open with your product — lead with the seller or shared context
- Use question-based approach, not statement-based
- No jargon: "daily payouts" not "factoring" or "receivables financing"
- One link max per message
- Tone: genuine, curious, peer-to-peer

Seller context:
- Name: {name}
- Business: {business_name}
- Category: {category}
- Marketplace: {marketplace}
- LinkedIn activity: {recent_posts_or_activity}
- Sequence step: {step_number}

Trigger: EventBridge every 2 hours, 7am–9pm ET Mon–Fri

Steps:
1. Read from AgentCore Memory (linkedin-sequences + cross-agent-leads):
   - Contacts needing next sequence action based on timing
   - Skip contacts already responded to email/voice (from cross-agent-leads)
   - Skip contacts in bdr_handoff or disqualified stage
2. For each contact:
   a. Read per-profile action counts from AgentCore Memory — AgentCore Policy
      guardrail will block the tool call if daily limit is reached
   b. Generate personalized message via Claude API (AgentCore Identity credential)
   c. Execute via AgentCore Browser (Nova Act → Playwright fallback):
      → Open AgentCore Browser session using persistent LinkedIn profile
        (AgentCore Identity manages session cookies — no manual loading)
      → Try Nova Act: connect to AgentCore Browser via CDP, execute task
      → If Nova Act escalates/errors: Playwright connects to AgentCore Browser
        via CDP, executes same task using CSS selectors
      → If both fail: log to dead-letter list in AgentCore Memory, skip contact today
   d. Write result to AgentCore Memory: advance sequence step, log executor used
   e. Call `hubspot_update_stage` tool to log activity

3. Inbox reply sweep (runs as part of same session):
   a. Open AgentCore Browser, run Nova Act inbox check → Playwright fallback
   b. For each new reply: classify intent via Claude API
   c. "interested": call `hubspot_create_task` tool (BDR alert),
      write to cross-agent-leads memory, update lifecycle → engaged
   d. "question": generate response via Claude + send via AgentCore Browser task

Deployment: Amazon Bedrock AgentCore
Trigger: Invoked by linkedin-outreach Lambda per queued action
Model: Nova 2 Lite (low-latency computer-use model)
Monitoring: CloudWatch — action success rate, escalation events, daily limit approach

Supported task types (passed as natural language instructions):

  Connection request:
    "Log into LinkedIn as {profile}. Navigate to this profile: {linkedin_url}.
     Send a connection request with the following note (under 300 characters): {note}.
     Confirm the request was sent successfully."

  Message to connection:
    "Log into LinkedIn as {profile}. Open the message thread with {contact_name}.
     Send the following message: {message_text}. Confirm delivery."

  Inbox check (reply sweep):
    "Log into LinkedIn as {profile}. Open the messaging inbox. Return a list of
     all conversations with new replies in the last 4 hours. For each, return:
     sender name, sender profile URL, and message text."

  Engagement (like/comment):
    "Log into LinkedIn as {profile}. Navigate to this post: {post_url}. Like it.
     Then leave this comment: {comment_text}."

Session management: Loads LinkedIn session cookie from Secrets Manager per profile.
Nova Act's step-by-step execution produces human-like timing — no explicit delays needed.

Triggers fallback when:
- Identity verification or security challenge screen appears
- "Connect" button not found on profile page (UI layout change)
- Message thread cannot be located
- Task exceeds 60 seconds
- Nova Act returns an escalation event for any reason

Execution: Playwright connects to AgentCore Browser via CDP (not a local Chrome instance)
Trigger: Invoked by the LinkedIn Agent's reasoning loop when Nova Act escalates or errors
Runtime: Python 3.12 + playwright (no headless Chromium layer needed — browser runs in AgentCore)
Proxy: Configured via AgentCore Browser proxy settings if needed (no external proxy service)

Session management: Load LinkedIn session cookie from Secrets Manager.
Add random delays between actions: 30–90 seconds (uniform distribution).
Randomize viewport (1200–1440px wide, 700–900px tall). Set realistic user-agent.

Connection request:
1. Navigate to {linkedin_url} via proxy
2. Wait for profile page to load, locate "Connect" button
   (selector: button[aria-label*="Connect"] — update if LinkedIn changes)
3. If "Add a note" option appears: click it, type {note}, click "Send"
4. Confirm via success toast or button state change
5. Return {success: true, executor: "playwright"}

Message to connection:
1. Navigate to linkedin.com/messaging/
2. Search for {contact_name} in conversation search
3. Click conversation, type {message_text} into compose box, send
4. Return {success: true, executor: "playwright"}

Inbox check (reply sweep):
1. Navigate to linkedin.com/messaging/?filter=unread
2. Scrape all visible conversation previews updated in last 4 hours
3. For each: extract sender name, profile URL (from href), message preview text
4. Return list of {sender_name, linkedin_url, message_text}

Engagement (like/comment):
1. Navigate to {post_url}
2. Click Like button (selector: button[aria-label="Like"])
3. Click comment box, type {comment_text}, submit
4. Return {success: true, executor: "playwright"}

On any Playwright failure: return {success: false, error: "<reason>", executor: "playwright"}
and push action back to dead-letter SQS for manual review.