All podcasts / Lenny's Podcast / Summary

Why LinkedIn is turning PMs into AI-powered "full stack builders" | Tomer Cohen (LinkedIn CPO)

2025-12-04 - 68 min - source - Read full transcript
Lenny Rachitsky (host)Tomer Cohen

Key insights

LinkedIn scrapped its APM pipeline and created a formal "Full Stack Builder" job title and career ladder, open to any function, that lets one builder take a product from idea to launch.
Cohen frames this as a structural response to a discontinuity: the skills required to do a given job will change by roughly 70% by 2030, and the market's fastest-growing jobs are turning over just as fast, so the old specialized product-development pipeline can no longer keep pace.
role-collapse
LinkedIn's product process didn't get complex because the underlying work is hard - it got complex because every step quietly grew its own specialist over time.
Cohen traces the chain explicitly: research expanded into 10-15 separate information sources, reviews multiplied (design, privacy, security), and each valid sub-step eventually needed a dedicated person, which turned process complexity into organizational complexity and finally into microspecialization - the actual reason a small feature can take multiple teams and sprints to ship.
role-collapse
Cohen keeps five traits as strictly human and is trying to automate everything else: vision, empathy, communication, creativity, and above all judgment.
He calls judgment - making high-quality decisions in complex, ambiguous situations - the single most important builder trait, more important than raw iteration speed, because it determines whether an organization can actually match its pace of response to the pace of change around it.
role-collapse
Off-the-shelf AI tools consistently fail to work on LinkedIn's stack without deep customization, whether for coding, design, or trust review.
Cohen says this was one of the program's biggest surprises: Cursor, Copilot, Windsurf, and Figma all required LinkedIn to build a custom integration layer against its own legacy code, design systems, and institutional context before the tools became useful - none worked well straight out of the box.
ai-agent-tooling
LinkedIn builds narrow, single-purpose internal agents scoped to specific institutional knowledge rather than one general-purpose assistant.
Named examples include a trust agent (built by the head of trust, catching privacy/scam vulnerabilities a years-old spec had missed), a growth agent trained on LinkedIn's historical funnels and tests, a research agent trained on member personas and support tickets, and an analyst agent that queries the LinkedIn graph directly - each owned by the relevant domain expert rather than a central AI team.
ai-agent-tooling
Giving an agent unrestricted access to a company's full knowledge base fails; curated "golden examples" are what actually make it work.
Cohen says LinkedIn initially gave an agent full drive access and it "failed miserably and hallucinates like crazy" because it couldn't weight importance or resolve conflicting opinions in the data. He compares the fix - painstakingly filtering good examples - to the weeks he spent manually curating what counted as a good LinkedIn feed post over a decade earlier.
ai-agent-tooling
The earliest and biggest AI investment went into coding because it was the easiest layer to automate; the underinvested layer was the idea-to-design front end.
LinkedIn's "maintenance agent" now resolves roughly half of all failed builds automatically, and a QA agent handles more. The Full Stack Builder push specifically targets the earlier, harder-to-automate stages - research, spec, design - which Cohen says is where a large share of a feature's ultimate quality is actually decided.
ai-agent-tooling
AI tooling amplified LinkedIn's best performers first, not its average or struggling employees.
Cohen reports that top talent is using the new tools the most and giving the most detailed feedback, which he ties to a general trait of high performers: a continuous, innate drive to get better at their craft, independent of whether change is mandated from the top.
career-reinvention-under-ai
Tools alone don't drive adoption - only a small minority (Cohen estimates ~5%) will pick up new tools voluntarily; the rest need deliberate change management.
Cohen's playbook, modeled on LinkedIn's earlier desktop-to-mobile transition, combines incentive changes (AI fluency now factors into performance calibration and hiring), visible pilot pods that prove the model works before a formal rollout, and actively celebrating individual success stories in all-hands meetings.
org-culture-for-ai
A LinkedIn UX researcher used the internal agent tooling to move directly into an open growth-PM role - a jump Cohen says wasn't a viable career path before.
Cohen offers this as the clearest evidence that full-stack mindset, not formal training or title, is now the real gate on role mobility; he frames the current moment as an unusually good time to cross between product, design, and engineering career tracks.
career-reinvention-under-ai

Books referenced

Media referenced

Companies

Techniques and frameworks

Summary

Tomer Cohen, LinkedIn's longtime chief product officer, used this conversation - recorded shortly before his departure after 14 years at the company - to lay out the "Full Stack Builder" model LinkedIn has been piloting internally. The premise is structural, not aspirational: Cohen cites LinkedIn's own labor-market data showing the skills required for a given job will change by roughly 70% by 2030, and that a majority of today's fastest-growing jobs didn't exist on the list a year ago. His conclusion is that the traditional product-development pipeline - research, spec, design review, code, launch, each staffed by an increasingly specialized function - has become too slow and too brittle for that pace of change, even though every individual step in that pipeline was added for a defensible reason.

The Full Stack Builder model collapses that stack back down. LinkedIn scrapped its Associate Product Manager (APM) program and replaced it with an Associate Product Builder (APB) track that trains new hires to code, design, and do product management before placing them into small, cross-functional "pods" that reassemble every quarter or so. There is now a formal "Full Stack Builder" job title and career ladder inside LinkedIn. Cohen is explicit that this isn't about eliminating specialization - some people, he says, genuinely don't want to be full-stack builders, and that's fine - but about shifting where humans spend their time: toward vision, empathy, communication, creativity, and especially judgment, while automating nearly everything else.

The tooling half of the conversation is the most concrete. Cohen repeatedly stresses that off-the-shelf AI tools "never" worked directly against LinkedIn's stack - not Cursor, not Copilot, not Figma - and that meaningful use required LinkedIn to re-architect its own platform to be AI-legible and to work in "alpha mode" with vendors to customize their tools against LinkedIn's codebase and design system. Rather than building one general-purpose assistant, LinkedIn built narrow, domain-owned agents: a trust agent (built by the head of trust) that catches privacy and scam vulnerabilities a years-old feature spec had missed; a growth agent trained on LinkedIn's historical funnels; a research agent trained on member personas and support-ticket history; and an analyst agent that queries LinkedIn's graph directly. The single biggest technical lesson, in Cohen's telling, was that dumping a company's full knowledge base into an agent backfires - it hallucinates and can't weight what matters - and that curating a small set of "golden examples" is what actually makes these tools useful, a discipline he compares to the manual work of defining what a "good" LinkedIn feed post looked like over a decade earlier.

The final third of the conversation covers adoption, which Cohen treats as harder than the technology itself. He estimates only around 5% of any organization will pick up new tools purely because they're available; everyone else needs explicit change management - incentive shifts (AI fluency now factors into LinkedIn's performance calibration), visible pilot pods that prove the model works before a company-wide rollout, and deliberately celebrated success stories, such as a LinkedIn UX researcher who used the new tooling to move directly into an open growth-PM role. Cohen frames the whole effort as a continuous process rather than a fixed end state, closing on his personal motto - "becoming is better than being" - as a lens for both the program and his own next chapter after leaving LinkedIn.

Notable Quotes

"It's not enough to give them the tools. You have to build the incentives programs, the motivation, the examples to how you do it." - Tomer Cohen

"We took every step and we expanded it to a lot of sub-steps... each one of those sub-steps actually has a valid reason to exist. But when you add a whole thing together, you're like, oh my God, this is why it takes multiple teams, multiple code bases, multiple sprints just to get out to launch." - Tomer Cohen

"It's not great to just give it access to your drive and say reason over all this knowledge base. It actually does a very poor job understanding importance of the past and putting weights on stuff." - Tomer Cohen

"If you're looking for a formal reorg or declaration to start building differently, you're waiting too long. Here's a permission for me to just not wait and just go." - Tomer Cohen

"Some people do not want to be full stack builders. And that's completely okay. Some people see themselves in specialization, and I think specialization has a place and a role." - Tomer Cohen