The head of engineering for Anthropic's Claude Code and Cowork (agent product for broader knowledge work) products describes a closed loop in which the same organization builds the frontier AI model, builds the coding product that runs on that model, and runs its own engineering…
Fiona Fung runs engineering and product management for Anthropic's Claude Code and Cowork tools, and on a recent Lenny's Newsletter episode she described a team that, in her telling, ships roughly 8x more code than it did before AI coding agents absorbed the routine work. Read past the productivity number, though, and the more interesting story is structural. Anthropic is the only organization in the world that simultaneously builds the frontier model, builds the coding product that runs on that model, and uses that product inside its own engineering org. That closed loop is the actual moat. The 8x is downstream of it.
In the interview, Fung frames her team's practice around Claude Code, Anthropic's AI coding assistant, and Cowork, the adjacent agent product aimed at broader knowledge work. The teams sit inside the same company that trains the underlying model, which means the feedback from internal engineering use flows back into the product roadmap on a cycle that competitors buying access to the same model cannot replicate. A team that is just a customer of Claude Code, even a sophisticated one, sees the model as a black box. Fung's team sees the model's failure modes as a release queue.
The other half of the story is what the loop does not solve. Fung is candid that blurring the line between engineer, reviewer, and manager introduces new costs: context-switching between human teammates and parallel agent runs, accountability for code no human fully read, and the management overhead of running a mixed human-and-agent org. Those problems are not unique to Anthropic, and they are not solved by adopting the same tools. They are open even for the team that built the tools.
This matters for the broader debate about AI-native engineering orgs. The Anthropic example is now cited in every internal conversation about what an AI-era software team should look like, and Fung's episode is one of the first long-form accounts from a senior engineering lead at a frontier lab about how that actually works day to day. The temptation is to copy the visible practices: agents in standups, blurred role definitions, an AI-first culture doc. But the visible practices are the surface. The 8x is the output of a vertical integration that only a model lab can run on itself.
The testable hypothesis for any other engineering leader is simple. If your organization does not build the model underneath its coding tool, the Anthropic playbook is a partially broken map. The standups will not hurt you. The role blurring will not hurt you. But the loop that makes the 8x real — model, product, internal practice, feeding back into the next model — is not for sale.
Fung's background underlines how unusual her vantage point is: roughly eleven years on Visual Studio and TypeScript at Microsoft, then a stretch at Meta where she says she started Facebook Marketplace, worked on Meta's first smart glasses and AR glasses effort, and led infrastructure, growth, integrity, and safety teams at Instagram, per the Lenny's Newsletter episode page for her interview with host Lenny Rachitsky. Twenty-five years of engineering, in other words, before she took on the Claude Code and Cowork organization.
What to watch next: whether Anthropic publishes the mechanisms behind its 8x claim with enough specificity that an outside team can stress-test the model, and whether any of the closed-loop advantages start to leak as competitors build tighter integrations with frontier model providers. The interesting question is not whether AI-pilled teams are faster. It is whether the speed comes from the tools, or from owning the stack the tools run on.