Brainforge Knowledge

MESSAGE_SEQUENCE_AGENT

9 min read

Message Sequence Agent

Purpose: AI-powered message sequence generation and approval system
Model: Vercel Lead Agent architecture adapted for GTM campaigns
Status: Reference Architecture
Last Updated: 2026-01-26

Overview

Message Sequence Agent is an AI-powered system for generating personalized, multi-step outreach sequences across different GTM campaigns. It uses durable workflows, AI-powered personalization, and human-in-the-loop Slack approval before sending.

This serves as a reference architecture to be adapted to your specific campaigns and workflows.

Key Features

✅ Immediate Response

Returns success to trigger upon campaign initiation

🔄 Durable Workflows

Uses workflow orchestration for reliable multi-step execution

🤖 Deep Research Agent

Conducts comprehensive research on prospects using AI SDK Agent class

📧 Sequence Generation

Generates personalized multi-step sequences with AI SDK generateObject

👤 Human-in-the-Loop

Sends to Slack for human approval before sending sequences

📊 Multi-Campaign Support

Supports multiple campaign types with different sequence strategies

Architecture

Campaign Trigger
     ↓
start(workflow) ← (Workflow DevKit)
     ↓
Research Agent ← (AI SDK Agent)
  - Prospect research
  - Company intelligence
  - Personalization signals
     ↓
Campaign Classifier ← (AI SDK generateObject)
  - Identify campaign type
  - Select appropriate sequence
  - Determine persona
     ↓
Sequence Generator ← (AI SDK generateObject)
  - Generate multi-step messages
  - Personalize each touchpoint
  - Apply campaign playbook
     ↓
Slack Approval ← (Slack Bolt + Vercel adapter)
  - Display sequence preview
  - Approve/Edit/Reject buttons
  - Edit in Slack modal
     ↓
Schedule & Send ← (On approval)
  - Queue messages
  - Track engagement
  - Update CRM

Tech Stack

ComponentTechnology
FrameworkNext.js 16
Durable ExecutionWorkflow DevKit
AIVercel AI SDK with AI Gateway
Human-in-the-LoopSlack Bolt + Vercel Slack Bolt adapter
ResearchExa.ai, Perplexity API
CRM IntegrationHubSpot API
DatabaseVercel Postgres (workflow state)

Campaign Types

1. Event Follow-Up

  • Trigger: Event attendance, booth visit, meeting log
  • Segments: Booth Visitor, Panelist, VIP Dinner, Meeting Log, General Attendee
  • Sequence: 3-5 touch points over 14 days
  • Personalization: Event context, conversation topics, engagement details

2. Mutual Introduction

  • Trigger: Intro request from mutual connection
  • Segments: Warm intro, cold intro with context
  • Sequence: 2-3 touch points over 7 days
  • Personalization: Mutual connection, referral context, shared interests

3. Partnership Outreach

  • Trigger: Partner rep engagement, co-sell opportunity
  • Segments: Cold rep, engaged rep, active deal support
  • Sequence: 4-6 touch points over 21 days
  • Personalization: Partner context, account research, value props

4. Product Launch

  • Trigger: New feature release, product update
  • Segments: Active users, churned users, prospects
  • Sequence: 3-4 touch points over 10 days
  • Personalization: Usage patterns, feature fit, use cases

5. Content Syndication

  • Trigger: New blog post, case study, whitepaper
  • Segments: By persona, by industry, by engagement history
  • Sequence: 2-3 touch points over 7 days
  • Personalization: Content relevance, industry fit, pain points

6. Re-engagement

  • Trigger: Inactive prospect, stalled deal, churned customer
  • Segments: By last engagement type, by stage
  • Sequence: 3-5 touch points over 21 days
  • Personalization: Last interaction, objections, new value

Workflow Steps

Step 1: Research Agent

Purpose: Gather intelligence for personalization

Inputs:

  • Prospect name, email, company
  • Campaign trigger context
  • Historical engagement data

AI Agent Tools:

  • searchWeb() - Web search via Exa.ai
  • queryKnowledgeBase() - Search internal docs
  • fetchCompanyData() - Company intelligence
  • analyzeLinkedIn() - LinkedIn profile analysis
  • getEngagementHistory() - Past interactions

Output:

{
  prospect: {
    name: string;
    title: string;
    company: string;
    linkedin: string;
  };
  intelligence: {
    companyInfo: string;
    recentNews: string[];
    painPoints: string[];
    competitors: string[];
  };
  personalization: {
    signals: string[];
    context: string;
    relevance: string;
  };
  engagementHistory: {
    lastContact: Date;
    interactions: number;
    stage: string;
  };
}

Step 2: Campaign Classifier

Purpose: Identify campaign type and select sequence strategy

Uses: AI SDK generateObject with schema

Output:

{
  campaignType: 'event-follow-up' | 'mutual-intro' | 'partnership' | 'product-launch' | 'content-syndication' | 're-engagement';
  segment: string; // Campaign-specific segment
  persona: 'executive' | 'practitioner' | 'champion' | 'blocker';
  sequenceLength: number; // 2-6 touchpoints
  priority: 'high' | 'medium' | 'low';
  reasoning: string;
}

Step 3: Sequence Generator

Purpose: Generate personalized multi-step sequence

Uses: AI SDK generateObject with campaign playbook context

Inputs:

  • Research report
  • Campaign classification
  • Campaign playbook (from memory)
  • Message templates

Output:

{
  sequence: [
    {
      step: 1;
      channel: 'email' | 'linkedin' | 'slack';
      timing: 'immediate' | 'day-2' | 'day-7' | 'day-14';
      subject: string;
      body: string;
      cta: string;
      personalizationNotes: string[];
    },
    // ... additional steps
  ];
  fallbackSequence: [...]; // If no response
  metadata: {
    generatedAt: Date;
    model: string;
    confidence: number;
  };
}

Step 4: Slack Approval

Purpose: Human review before sending

Slack Message Format (Block Kit):

{
  "blocks": [
    {
      "type": "header",
      "text": {
        "type": "plain_text",
        "text": "📧 New Sequence: [Campaign Type]"
      }
    },
    {
      "type": "section",
      "fields": [
        {"type": "mrkdwn", "text": "*Prospect:*\n[Name] @ [Company]"},
        {"type": "mrkdwn", "text": "*Campaign:*\n[Type]"},
        {"type": "mrkdwn", "text": "*Segment:*\n[Segment]"},
        {"type": "mrkdwn", "text": "*Steps:*\n[N] touchpoints"}
      ]
    },
    {
      "type": "divider"
    },
    {
      "type": "section",
      "text": {
        "type": "mrkdwn",
        "text": "*Step 1 (Immediate):*\n```\nSubject: [Subject]\n\n[Body Preview]\n```"
      }
    },
    {
      "type": "actions",
      "elements": [
        {
          "type": "button",
          "text": {"type": "plain_text", "text": "✅ Approve & Send"},
          "style": "primary",
          "value": "approve"
        },
        {
          "type": "button",
          "text": {"type": "plain_text", "text": "✏️ Edit"},
          "value": "edit"
        },
        {
          "type": "button",
          "text": {"type": "plain_text", "text": "❌ Reject"},
          "style": "danger",
          "value": "reject"
        }
      ]
    }
  ]
}

Edit Modal: Opens Slack modal to modify sequence before sending

Step 5: Schedule & Send

Purpose: Queue and send approved sequence

Actions:

  1. Store sequence in CRM (HubSpot)
  2. Schedule messages based on timing
  3. Set up engagement tracking
  4. Monitor responses for triggers
  5. Update workflow state

Project Structure

message-sequence-agent/
├── app/
│   ├── api/
│   │   ├── trigger/          # Campaign trigger endpoint
│   │   ├── slack/            # Slack webhook handler
│   │   └── webhook/          # CRM/external webhooks
│   └── page.tsx              # Dashboard
├── lib/
│   ├── services/
│   │   ├── research.ts       # Research agent
│   │   ├── classifier.ts     # Campaign classification
│   │   ├── generator.ts      # Sequence generation
│   │   ├── sender.ts         # Email/message sending
│   │   └── crm.ts            # CRM integration
│   ├── slack.ts              # Slack integration
│   ├── types.ts              # TypeScript schemas
│   └── playbooks/            # Campaign playbooks (imported)
│       ├── event-follow-up.ts
│       ├── mutual-intro.ts
│       ├── partnership.ts
│       └── ...
├── workflows/
│   └── sequence/
│       ├── index.ts          # Main workflow
│       └── steps.ts          # Workflow steps
└── memory/                   # Campaign playbooks (markdown)
    ├── event-follow-up-playbook.md
    ├── mutual-intro-playbook.md
    ├── partnership-playbook.md
    └── ...

Configuration

Environment Variables

# Vercel AI Gateway
AI_GATEWAY_API_KEY=xxx
 
# Slack Bot
SLACK_BOT_TOKEN=xoxb-xxx
SLACK_SIGNING_SECRET=xxx
SLACK_CHANNEL_ID=xxx
 
# Research APIs
EXA_API_KEY=xxx
PERPLEXITY_API_KEY=xxx
 
# CRM
HUBSPOT_API_KEY=xxx
 
# Database
POSTGRES_URL=xxx

Key AI Prompts

Research Agent System Prompt

const RESEARCH_AGENT_PROMPT = `
You are a GTM research agent for Brainforge, an embedded data & analytics consultancy.
 
Your job is to research prospects and identify personalization signals for outreach.
 
Focus on:
1. Company intelligence (stage, funding, tech stack, recent news)
2. Prospect background (role, tenure, previous companies)
3. Pain points (data infrastructure, analytics, team gaps)
4. Personalization opportunities (mutual connections, shared interests, relevant content)
 
Be thorough but concise. Prioritize actionable insights over general information.
`;

Campaign Classifier Prompt

const CLASSIFIER_PROMPT = `
Classify this campaign based on the trigger context and research report.
 
Campaign Types:
- event-follow-up: Post-event outreach (conferences, meetings, dinners)
- mutual-intro: Warm introduction from mutual connection
- partnership: Partner rep engagement or co-sell
- product-launch: New feature announcement
- content-syndication: Blog/case study sharing
- re-engagement: Reactivate inactive prospect/customer
 
Segments vary by campaign type. Use the research report to determine the most relevant segment.
 
Persona Types:
- executive: C-level, VPs (focus on ROI, outcomes, strategic value)
- practitioner: Engineers, analysts (focus on technical depth, use cases)
- champion: Internal advocate (focus on enablement, success stories)
- blocker: Skeptical or resistant (focus on risk mitigation, proof points)
 
Select sequence length (2-6 steps) based on:
- Engagement history (more history = shorter sequence)
- Persona (executives prefer shorter)
- Campaign type (events need faster follow-up)
`;

Sequence Generator Prompt

const GENERATOR_PROMPT = `
Generate a personalized, multi-step outreach sequence using the campaign playbook.
 
Guidelines:
1. Apply playbook structure and messaging principles
2. Use research insights for personalization (not generic)
3. Match persona tone and style
4. Include specific CTAs for each step
5. Follow timing recommendations from playbook
6. Add personalization notes for human reviewer
 
Remember: Not "Great to meet you!" — narrative-based messaging with specific context.
`;

Extensibility

Adding New Campaign Types

  1. Create playbook markdown in memory/

    # [Campaign Name] Playbook
 
## Overview
## Segmentation Strategy
## Persona-Based Messaging
## Multi-Step Sequences
## Examples

  2. Add campaign type to lib/types.ts

    export type CampaignType = 
  | 'event-follow-up'
  | 'my-new-campaign';

  3. Import playbook to lib/playbooks/my-new-campaign.ts

    import playbookMd from '@/memory/my-new-campaign-playbook.md';
export const playbookContext = playbookMd;

  4. Update classifier to recognize new campaign type

  5. Test with sample triggers

Adding New Sequence Steps

Modify the sequence schema in lib/types.ts:

export const sequenceStepSchema = z.object({
  step: z.number(),
  channel: z.enum(['email', 'linkedin', 'slack', 'phone']),
  timing: z.string(),
  subject: z.string().optional(),
  body: z.string(),
  cta: z.string(),
  attachments: z.array(z.string()).optional(),
  personalizationNotes: z.array(z.string()),
});

Adding New Research Tools

Add tools to the research agent in lib/services/research.ts:

const agent = new Agent({
  model: openai('gpt-4'),
  tools: {
    searchWeb,
    queryKnowledgeBase,
    fetchCompanyData,
    analyzeLinkedIn,
    getEngagementHistory,
    // Add new tool here
    myCustomTool: tool({
      description: 'Description for AI',
      parameters: z.object({ ... }),
      execute: async ({ params }) => {
        // Implementation
      },
    }),
  },
});

Human-in-the-Loop Best Practices

Approval Workflow

  1. Always show full context in Slack message

    • Prospect details
    • Campaign type and reasoning
    • Full sequence preview (or first 2 steps)

  2. Make editing easy

    • Open modal with editable fields
    • Preserve formatting
    • Save edits before sending

  3. Track approval metrics

    • Approval rate by campaign type
    • Common edits/rejections
    • Use feedback to improve prompts

  4. Timeout handling

    • Auto-reject after 24 hours
    • Send reminder after 4 hours
    • Allow manual trigger later

Quality Control

  • Review rejection patterns to improve generation
  • A/B test different prompts and templates
  • Monitor response rates by campaign and segment
  • Iterate on playbooks based on what works

Success Metrics

Generation Quality

  • Approval Rate: Target 80%+ on first generation
  • Edit Rate: <30% require significant edits
  • Rejection Rate: <10% rejected outright

Sequence Performance

  • Open Rate: Target 40-50% on initial email
  • Response Rate: Target 15-20% across sequence
  • Meeting Booked: Target 8-12% conversion
  • Opt-out Rate: <2%

Efficiency

  • Time to Generate: <2 minutes end-to-end
  • Human Review Time: <3 minutes per sequence
  • Sequences/Week: Target 50-100 sequences

License

Reference architecture for internal use.

Graph View

Backlinks