Brainforge Knowledge

VERCEL_LEAD_AGENT_LEARNINGS

12 min read

Learnings from Vercel’s Lead Agent (Drew Brevik Interview)

Context: How Vercel built a lead agent that drove $2M revenue and reduced SDR team from 20 to 2 people Key Insight: Built in a weekend, took 1 week dev, 5 weeks to convince. The real blocker was buy-in, not building.

🎯 Key Takeaways for Brainforge

1. Start Small, Prove Value Fast

What Drew Did:

  • Built initial prototype in a weekend
  • Finished dev work in 1 week
  • Spent 5 weeks getting buy-in (the real blocker)

Lesson for You:

  • Don’t wait for perfect planning - build something quick
  • Expect resistance, budget time for convincing stakeholders
  • Data speaks louder than promises - run in “shadow mode” first

Application:

Your First Agent: Pick one repetitive task
1. Build MVP in 1-2 days
2. Test on historical data (back testing)
3. Run in production (shadow mode) for 2-4 weeks
4. Use data to get buy-in
5. Deploy when proven

🔍 Finding the Right Agent Opportunity

Drew’s Framework: Two Types of Agents

1. Efficiency Agents (What you’re already doing)

  • Pattern: Boring, repetitive tasks done a lot (100+ times/day)
  • Example: SDR lead qualification (15-20 min each, all day long)
  • Human in Loop: YES (keep humans, they’re one-way doors)
  • Principle: “You’re already paying humans to do it, so business cares”

2. “Should” Agents (What you should be doing but aren’t)

  • Pattern: Things that aren’t getting done (not deemed worth paying for)
  • Examples:
    • Postmortems on lost deals
    • Extracting objections from sales calls
    • Working the bottom 2/3 of lead lists
  • Human in Loop: NO (can fire by default)
  • Principle: “Stuff that hasn’t been deemed worth minimum wage”

How to Identify Agent Opportunities

The Venn Diagram:

AI Capability (Repetitive, Pattern-Based)
    ∩
High Frequency (100+ times/day)
    ∩
Tedious/Unfulfilling Work
    =
Perfect Agent Opportunity

Questions to Ask:

  1. Is it repetitive? (Same task over and over)
  2. Is it high frequency? (Happening 100+ times/day/week)
  3. Is it tedious? (No one wants to do it)
  4. Do we have historical data? (For back testing)

If all YES → Great candidate for an agent

🏗️ The Build Process (Drew’s Method)

Step 1: Shadow Your Best People

What Drew Did:

  • Sat with best SDR at Vercel
  • Watched their actual process (not what they say they do)
  • Learned they check LinkedIn, look up company, check Versell account usage
  • Documented the “revealed preference” vs “stated preference”

Lesson: Revealed Preference > Stated Preference

  • People say they want walkable cities → They live in suburbs and drive SUVs
  • People say they follow a process → They actually do shortcuts/optimizations

For You:

  1. Find your A+ performers in each function
  2. Shadow them doing the actual work
  3. Document what they REALLY do (not what the process says)
  4. Capture their tricks/optimizations
  5. Build agents that replicate the best behavior

Questions to Ask While Shadowing:

  • “What are you checking first?”
  • “How do you decide if it’s worth pursuing?”
  • “What shortcuts do you take?”
  • “What do you always check that others skip?”

Step 2: Build Throwaway UI for Prompt Iteration

What Drew Did:

  • Built quick UI in Cursor/Claude Code
  • Pulled 3 months of historical leads from Salesforce
  • Iterated on prompt until agent agreed with human decisions
  • Used back testing (evaluations) to refine

The Process:

1. Load historical data (3 months of leads)
2. Run agent on each historical lead
3. Compare agent decision vs. human decision
4. For disagreements:
   - Was human wrong? → Update your mental model
   - Was AI wrong? → Update prompt
5. Iterate until "don't miss a single free throw"

Key Insight: Prompt Structure Matters

  • Put reasoning FIRST, decision LAST
  • This forces “System 2” thinking (slow, analytical) vs “System 1” (snap judgment)
  • Same principle as “Thinking Fast and Slow” - applies to LLMs too

Prompt Structure:

BAD:
"Classify this lead as qualified or not qualified. [reasons below]"

GOOD:
"Analyze this lead. First, list your reasoning:
1. [Reason 1]
2. [Reason 2]
3. [Reason 3]

Based on this reasoning, your decision: [qualified/not qualified]"

Step 3: Get Into Production (Shadow Mode)

What Drew Did:

  • Connected to production via webhook from contact form
  • Ran agent in parallel to existing process
  • Didn’t send emails yet - just generated them
  • Humans could see what agent would do vs. what they actually did
  • Collected metrics to prove value

The Setup:

Contact Form → Enrichment (Clearbit) → CRM Enrichment → Lead Agent → Slack Notification

Lead Agent Output:
- Research report (deep research on company)
- Qualification decision
- Reasoning
- Draft email (if qualified)
- Human approves → Sends email

Why Shadow Mode Works:

  • Proves agent works with real data (not just historical)
  • Lets you catch edge cases before full deployment
  • Builds trust through visibility
  • Gives you data for buy-in conversations

Step 4: Prove Value, Get Buy-In

Drew’s Approach:

  1. Got buy-in from ICs (individual contributors) first
  2. Partnered with functional leader (Nick Gage, SDR manager)
  3. Data analyst found 100% efficiency improvement (8 touch points → 4)
  4. Present to leadership with data, not promises

Metrics That Matter:

  • Efficiency: Touch points per lead (8 → 4 = 100% improvement)
  • Time savings: Hours per day/week saved
  • Quality: Agreement rate with humans (should be 90%+)
  • Business impact: Revenue, pipeline, headcount reduction

The Pitch:

  • Not: “AI can help us save time”
  • Yes: “We saved 100% on touch points, freed up 18 SDRs, they’re now doing outbound (higher value), and we’re generating $2M incremental revenue”

💻 Technical Implementation

Tech Stack (What Drew Used)

Core:

  • Next.js app (frontend + API routes)
  • Vercel (deployment)
  • Workflow devkit (long-running backend tasks)

AI:

  • AI SDK (abstraction layer for multiple providers)
  • AI Gateway (load balancing, fallback logic)
  • Multiple models (10-12 models across different agents)

Integrations:

  • Salesforce (CRM data)
  • Clearbit (enrichment)
  • Exa.ai (web search)
  • Slack (notifications via Bolt)
  • Turbopuffer (vector database for knowledge base)

Key Insight: AI Gateway is Critical

  • Anthropic goes down → Falls back to other Claude providers
  • All Claude down → Falls back to GPT
  • Prevents business disruption
  • Single point of failure is dangerous

The Agent Structure (Surprisingly Simple)

What an Agent Actually Is:

const agent = new Agent({
  model: "gpt-4",
  system: "[your prompt here]",
  tools: [search, queryKnowledgeBase, fetchURL],
  stopWhen: (result) => result.toolCalls >= 20
});

That’s it. An agent is:

  1. Pick a model
  2. Write a prompt
  3. Give it tools
  4. Set stop criteria

The hard part is NOT the code - it’s:

  • Writing a good prompt
  • Providing quality data
  • Integrating with systems of record
  • Handling edge cases

Knowledge Base Architecture

Drew’s Setup (Turbopuffer):

  • Account-level knowledge base (one collection per account in Salesforce)
  • Documents are markdown representations of:
    • Every sales call
    • Every email
    • Internal Slack communications
    • External Slack communications
  • All primary sources embedded
  • All agents can query it

Why This Matters:

  • Prevents doing same research over and over across agents
  • Each agent doesn’t need to rebuild the wheel
  • Centralized, up-to-date knowledge
  • Cost-effective (Turbopuffer is $64/month vs. Pinecone much higher)

For Brainforge:

  • Your memory/ structure is similar concept
  • But consider vector database when you hit scale
  • Store account-level knowledge (clients, prospects)
  • Embed meeting transcripts, emails, project docs

🎯 Choosing Your First Agent

Drew’s Rule of Thumb

“Whatever the biggest growth/sales/marketing problem you have in your business, that’s a good spot to start because you’re going to be passionate about fixing it.”

Don’t:

  • Pick a small, unimportant project
  • Copy what worked for someone else (might not be 10x for you)
  • Build something that happens once a day

Do:

  • Pick your biggest pain point
  • Something you’re passionate about solving
  • High frequency (happening a lot)
  • Has historical data for testing

Candidate Evaluation Framework

Score each potential agent opportunity:

CriteriaScore (1-5)WeightTotal
Frequency (happens often?)3x
Business impact (revenue/time saved?)3x
Tedious/automate-able?2x
Historical data available?2x
Passion to solve it?1x

Highest score wins.

📊 The Prompt Refinement Process

Initial Prompt Development

Drew’s Method:

  1. Use your own taste/knowledge of business
  2. Shadow best people, incorporate their process
  3. Start with case-by-case analysis
  4. Build up list of edge cases

Iterative Refinement

The Loop:

1. Find disagreement (AI vs. human)
2. Determine: Who's wrong?
   - AI wrong → Add to prompt constraints
   - Human wrong → Learn from it, update mental model
3. Prompt getting unruly? → Use AI to help format it better
4. Repeat until 95%+ agreement

Key Insight: AI is Eager to Please

  • Most of Drew’s prompt is telling AI “NO, you don’t want this lead”
  • AI naturally tries to make you happy
  • You have to convince it that actually, you DON’T want this “good outcome”

Example Prompt Structure:

System: "You are a lead qualification agent. Your job is to QUALIFY leads, 
not to accept every lead. Most leads are NOT qualified. You should be 
REJECTING more than you accept.

Reasons to REJECT a lead:
1. Spam/tester (90% of contact forms)
2. Not decision maker
3. No budget
4. Not a good fit

Only accept if:
1. Clear decision maker
2. Budget confirmed
3. Good fit for product
4. Real business use case

[detailed qualification criteria...]

Reasoning first, then decision."

Prompt Versioning

Critical Practice:

  • Version your prompts (v1, v2, v3)
  • Store in source control (Git)
  • Tag which version was used for which leads
  • Makes it easy to:
    • Roll back if new version worse
    • A/B test different prompts
    • Debug issues later

🤝 Getting Buy-In and Adoption

The 5-Week Buy-In Process

Week 1-2: Shadow Mode in Production

  • Run agent in parallel
  • Don’t actually use its output yet
  • Collect metrics

Week 3: IC Buy-In

  • Get individual contributors (SDRs) to say “this is pretty good”
  • They’re on the ground, their opinion matters
  • They’ll become champions

Week 4: Functional Leader Partnership

  • Partner with functional leader (SDR manager)
  • Get them to co-sign
  • They have stake in success

Week 5: Leadership Presentation

  • Present data, not promises
  • Show metrics (efficiency improvement, time saved)
  • Let data speak

Key Insight: Most blockers are human, not technical.

Overcoming Resistance

Common Fears:

  • “Hallucinations will hurt us”
  • “AI will make mistakes”
  • “We’ll lose control”

Drew’s Approach:

  • Address fears head-on
  • Human in loop (for efficiency agents)
  • Show data on accuracy
  • Start with low-risk use cases

For “Should” Agents:

  • Less resistance (you’re not taking away someone’s job)
  • Can fire by default (no human needed)
  • They do stuff that wasn’t getting done anyway

💡 Practical Applications for Brainforge

Immediate Opportunities (Based on Your Vault)

1. Sales Proposal Agent (Efficiency)

  • What: Auto-generate proposals/SOWs from discovery notes
  • Frequency: Every new deal (high frequency)
  • Current Process: You manually create from gtm/sales/hedra-sow.md templates
  • Agent Could:
    • Read discovery notes from lmnt/meeting_notes/
    • Pull template from memory/examples/proposals/
    • Generate personalized SOW
    • Human reviews and sends

2. Meeting Summary Agent (“Should”)

  • What: Auto-summarize every meeting transcript
  • Frequency: Multiple meetings per week
  • Current State: Manual summaries in meeting/summary/
  • Agent Could:
    • Read transcript from meeting/transcripts/
    • Extract key decisions, action items, insights
    • Store in appropriate memory/ or experiences/ location
    • Update relevant knowledge files

3. Customer Research Agent (Efficiency)

  • What: Deep research on prospects before calls
  • Frequency: Every new prospect/client
  • Current Process: Manual research (LinkedIn, company, etc.)
  • Agent Could:
    • Research company (like Vercel’s lead agent)
    • Pull relevant case studies from memory/customers/
    • Generate briefing doc
    • Human uses for call prep

4. Pattern Extraction Agent (“Should”)

  • What: Automatically extract patterns from completed work
  • Frequency: After every project/deliverable
  • Current State: Manual /learn process
  • Agent Could:
    • Read completed project from experiences/
    • Extract what worked/patterns
    • Update memory/patterns/
    • Increase pattern confidence scores

Build Order Recommendation

Week 1-2: Meeting Summary Agent (Easiest win, builds knowledge base)

  • Low risk (“should” agent)
  • High value (captures knowledge automatically)
  • Builds foundation for other agents

Week 3-4: Customer Research Agent (Efficiency, high frequency)

  • Shadow your best sales calls
  • Document research process
  • Build agent that replicates it

Week 5-6: Proposal Agent (Efficiency, high impact)

  • Start with template-based generation
  • Refine with feedback
  • Scale to full proposals

🚨 Common Pitfalls (What Drew Warned About)

1. Garbage In, Garbage Out

“Your agents will only be as good as the data you give them.”

Solution:

  • Invest in data quality
  • Clean up your CRM
  • Maintain good knowledge base
  • Regular audits of agent inputs

2. Copy-Paste Failure

“You can’t just steal what was a 10x for me. It might not be a 10x for you.”

Solution:

  • Understand YOUR business problems
  • Adapt frameworks to YOUR context
  • Test what works for YOUR team

3. Starting Too Small

“Don’t automate something that happens once or twice a day.”

Solution:

  • Focus on high-frequency tasks
  • Look for 100+ times/day or week
  • Otherwise, ROI doesn’t justify effort

4. Ignoring the Human Side

“1 week to build, 5 weeks to convince.”

Solution:

  • Budget time for buy-in
  • Partner with functional leaders
  • Get IC champions early
  • Let data do the talking

5. No Human in Loop (for Efficiency Agents)

“Efficiency agents keep humans in loop. They’re one-way doors.”

Solution:

  • Always have human review for efficiency agents
  • Only let “should” agents fire by default
  • Build trust before full automation

📋 Action Items for Brainforge

This Week

  • Identify your #1 repetitive task (what’s done 100+ times/week?)
  • Shadow your best person doing that task
  • Document the actual process (revealed preference)
  • Pull 3 months of historical data for back testing

This Month

  • Build MVP agent (throwaway UI)
  • Iterate on prompt until 90%+ agreement with humans
  • Get into production (shadow mode)
  • Start collecting metrics

Next Month

  • Present data to leadership
  • Get buy-in for full deployment
  • Start planning next agent

🎓 Key Quotes to Remember

On Choosing What to Build:

“Whatever the biggest growth/sales/marketing problem you have in your business, that’s a good spot to start because you’re going to be passionate about fixing it.”

On the Build Process:

“These projects can start small but they can grow into things that are big. So yeah, just get started, focus on a problem that actually matters to your business.”

On Data Quality:

“Your agents will only be as good as the data you give them. The crap in crap out thing is totally true, especially with AI.”

On Getting Buy-In:

“Really partner closely with your functional leads. There’s a lot of good in letting them prototype. Get your functional leaders a v0 seat. Have them sit down and think about ideas and build stuff themselves.”

On Prompt Structure:

“If you make it tell its reasons first and then give you the final answer at the end, it’s actually smarter. It’s just the way math works. It’s wild.”

Bottom Line: Start small, prove value fast, let data drive adoption. The hard part isn’t building - it’s finding the right opportunity and getting people to use it.

Graph View

Backlinks