Skip to content
Lenny's 🏦 Vault

Curated Data Improves AI Agent Performance

by Tomer Cohen on December 4, 2025

LinkedIn's Full Stack Builder Model: Transforming Product Development with AI

LinkedIn's strategic shift to a "Full Stack Builder" model demonstrates how AI can fundamentally transform product development processes at scale. Under CPO Tomer Cohen's leadership, LinkedIn is pioneering a new approach that collapses traditional organizational silos and empowers builders to take ideas from concept to launch with AI assistance.

Situation

  • Accelerating change: LinkedIn observed that skills required for jobs will change by 70% by 2030, and 70% of today's fastest-growing jobs weren't on the list a year ago
  • Organizational complexity: Over time, LinkedIn's product development process had become increasingly complex with specialized roles, multiple handoffs, and lengthy processes
  • Competitive necessity: The "time constant of change" had become greater than the "time constant of response" - change was happening faster than the organization could adapt
  • Legacy constraints: As a mature company with established systems, LinkedIn couldn't simply adopt off-the-shelf AI tools or completely reinvent overnight

Actions

Platform Transformation

  • Rearchitected core platforms to make them "AI-ready" so AI could reason over their codebase
  • Created composable UI components with server-side support specifically designed for AI interaction
  • Customized third-party tools like Figma, Copilot, and Cursor to work with LinkedIn's unique systems

AI Agent Development

  • Built specialized, purpose-specific AI agents rather than a single general-purpose agent:
    • Trust Agent: Evaluates potential vulnerabilities and harm vectors in product ideas
    • Growth Agent: Analyzes growth opportunities based on LinkedIn's unique loops and funnels
    • Research Agent: Trained on member personas and historical research data
    • Analyst Agent: Enables natural language querying of LinkedIn's data graph
    • Maintenance Agent: Automatically fixes failed builds (now handling ~50% of cases)

Knowledge Curation for AI

  • Critical insight: Initially gave agents access to all company knowledge, which "failed miserably"
  • Carefully selected and cleaned high-quality examples for training, similar to how they had curated examples for LinkedIn's feed algorithm years earlier
  • Focused on providing relevant context rather than overwhelming agents with all available information

Cultural Transformation

  • Created a formal "Full Stack Builder" career path that anyone from any function could pursue
  • Replaced their APM program with an "Associate Product Builder" program
  • Formed cross-functional "pods" of builders focused on missions rather than functional specialties
  • Updated performance reviews to evaluate AI fluency and cross-functional capabilities
  • Celebrated early wins and showcased examples of successful transitions (like a researcher becoming a growth PM)

Results

Early Outcomes

  • Teams are saving hours of work per week through AI assistance
  • Top performers are adopting tools most enthusiastically and providing valuable feedback
  • Designers are pushing code directly, and PMs are building their own dashboards
  • Cross-functional mobility has increased, with people transitioning between traditionally separate roles

Implementation Challenges

  • Off-the-shelf AI tools never worked without significant customization
  • Different teams gravitated toward different tools, creating standardization challenges
  • Some employees prefer specialization and don't want to become full-stack builders
  • Initial attempts to give AI access to all company knowledge led to hallucinations and poor results

Key Lessons

AI Implementation Strategy

  • Curate training data carefully: Don't just give AI access to your entire knowledge base. Select high-quality examples and clean the data, as LinkedIn learned from both their feed algorithm and AI agent development.
  • Build for specific use cases: Purpose-built agents for specific functions (trust, growth, research) outperform general-purpose solutions.
  • Customize, don't just adopt: Off-the-shelf AI tools rarely work without significant customization to your specific systems and context.

Organizational Transformation

  • Change management is critical: "It's not enough to give them the tools... you have to build the incentives, programs, motivation, and examples." Without this, only about 5% of employees will adopt.
  • Focus on mindset, not titles: "I could care less about your title. I care about how you work... changing your mindset to a full stack mindset is what I'm looking for."
  • Align individual and organizational incentives: Show how AI adoption benefits both the company's agility and individual career development.

Product Development Philosophy

  • Emphasize human judgment: Automate everything except vision, empathy, communication, creativity, and judgment - the latter being most critical.
  • Smaller, mission-focused teams: Structure as nimble "pods" focused on problems rather than large functional teams.
  • Becoming over being: Embrace continuous improvement rather than fixed states - "becoming is better than being."
View full episode