Tomer Cohen on the Full Stack Builder Model, AI-Native Teams, and LinkedIn’s Internal Agents

Tomer Cohen, outgoing Chief Product Officer at LinkedIn (14 years), in conversation with Lenny Rachitsky. The episode covers LinkedIn’s transition to a Full Stack Builder model — any-function employees taking ideas to market — the four internal agents LinkedIn built, the change management required to make adoption real, and reflections on Tomer’s departure.

Key ideas

Organisational complexity is the real enemy, not the technology. Every step in product development has expanded: research became 15 sources, shipping requires design + privacy + security reviews. Each sub-step is rational; together they produce microspecialisation and long cycle times. AI makes it possible to collapse the stack back down — but that requires a deliberate choice to do so.
Full Stack Builders run three parallel investments: platform, tools, culture. Platform means rearchitecting code so AI can reason over it. Tools means building specialised agents (trust, growth, research, analyst) — none of which work off the shelf. Culture means change management: performance reviews, visible wins, and peer examples. Culture is the constraint; technology is the prerequisite.
Top talent adopts first. LinkedIn’s early-access pods showed that the highest performers leaned in hardest. This is consistent with how every major platform shift (desktop to mobile) has played out. The implication: seeding the programme with your strongest people is not elitist — it is the correct adoption strategy.
Agents require curated knowledge, not raw access. Giving an agent access to a Google Drive “failed miserably.” The trust agent, growth agent, and research agent all required months of building gold-example corpora — curated samples of what good looks like — before they produced useful output.
“Becoming is better than being.” Tomer’s guiding motto. The Full Stack Builder is not a destination (a title) but a mindset (continuous iteration across domains). If you are waiting for a formal reorg to start building differently, you are waiting too long.

The Full Stack Builder model

The model has a single goal: empower any builder to take an idea to market, regardless of role or team.

LinkedIn defines the human contribution as five traits: vision (compelling sense of the future), empathy (profound understanding of unmet needs), communication (aligning others), creativity (possibilities beyond the obvious), judgment (high-quality decisions in ambiguity). Everything else is being automated.

The organisational form is pods — small assembled groups that tackle a problem for a quarter and then re-form. The analogy Tomer uses is Navy SEALs: cross-trained, mission-focused, small, and re-deployable.

The formal title Full Stack Builder now exists at LinkedIn. The APM programme ended; the APB (Associate Product Builder) programme replaced it in January 2026, training graduates in coding, design, and product management before they join pods.

The four internal agents

Agent	Domain	What it does
Trust agent	Trust & safety	Scans specs for vulnerability and harm vectors unique to LinkedIn
Growth agent	Product growth	Critiques ideas against LinkedIn’s historical loops and experiment results
Research agent	User research	Evaluates specs against member personas, past studies, and support tickets
Analyst agent	Data	Queries the LinkedIn graph without requiring SQL or data science round-trips

None of these came off the shelf. LinkedIn worked directly with vendors (including Copilot Enterprise, ChatGPT Enterprise) to customise. Building the orchestrating layer that routes between agents — so trust and growth agents collaborate, not just sequence — is ongoing work.

A fifth agent, the product jammer, orchestrates LinkedIn’s product-jam process and calls the other agents as needed.

Change management

“It’s not enough to give them the tools to use it. You have to build the incentives programmes, the motivation, the examples to how you do it.” — Tomer Cohen

Mechanisms that worked at LinkedIn:

Performance reviews now incorporate AI fluency and agency
360 reviews for Tomer’s direct reports included ratings from other functions (designers rating PMs, etc.)
Pod access as FOMO — limited early access with feedback requirements, not open roll-out
Celebrating wins: a user researcher became a growth PM using the tools; an head of partnerships built a developer portal himself
APB programme signals to the whole organisation that this is the direction

What failed: giving agents access to all company data with no curation. Expecting tools to work off the shelf.

Who is doing this elsewhere