What Anthropic's Engineering Leader Says About Actually Using AI to Ship Code

Who Is Talking and Why It's Worth Hearing

Most commentary on AI in engineering comes from people selling something or from people who have never run a team. This interview is different. Fiona Fung leads the teams behind Claude Code and Cowork at Anthropic, overseeing engineering and product for those products. Before that she spent 11 years at Microsoft building Visual Studio and TypeScript, then moved to Meta, where she worked on VR and AR glasses, started Facebook Marketplace, and led Instagram infrastructure, growth, and safety teams.

That's more than 25 years of shipping real software. So when she describes how AI is changing the way her teams work, it's coming from someone who has managed the messy reality of large engineering organizations, not just the demo.

The conversation covers specific ways her team uses AI, how they plan work, what she has learned running teams that ship far more code, which roles she expects AI to transform next, and the problems that come with all of this. That last part is the useful bit, because the downsides usually get skipped.

What the Conversation Actually Covers

A few themes stand out for anyone building automated or AI-assisted workflows.

The first is throughput. Fung discusses running teams that are shipping roughly 8x more code. That's a striking number, and it's worth being clear about what it does and doesn't mean. Shipping more code is not the same as shipping more value, and she frames it as something her teams had to learn to manage rather than a free win. When output goes up that fast, the bottleneck moves. Review, planning, coordination, and deciding what to build all become the constraint instead of the typing.

The second theme is how planning changes. If individual engineers can produce far more with AI assistance, the work of deciding direction and breaking down problems gets more important, not less. The interview goes into how her teams approach planning in that environment. For anyone automating their own work, this is the honest takeaway: the tools accelerate execution, so the human effort shifts up the stack toward judgment and scoping.

The third theme is the human cost. Fung raises what she calls an emerging context-switching and loneliness problem for engineers. When a single person is directing several AI agents at once, the nature of the work changes. You spend more time coordinating and reviewing and less time in the shared, collaborative flow that traditional engineering involved. That's a real observation about a real side effect, and it's rare to hear a leader name it plainly instead of pretending the transition is frictionless.

She also talks about which roles AI will transform next and what keeps her up at night. Rather than paraphrase claims that aren't spelled out in the summary, the honest thing to say is that she treats this as an open question with genuine risks, not a solved problem.

Why This Matters If You're Building Automations

The practical lesson here transfers well beyond a large AI lab. If you run a small team or automate your own workflows, the pattern Fung describes is the one you should plan for.

When AI removes the effort of execution, three things happen. Your output can climb sharply. Your bottleneck moves to review and decision-making. And your process for deciding what to work on becomes the thing that determines whether all that extra output is worth anything. If you've ever wired up a pipeline in n8n, Make, or Zapier that suddenly let you produce ten times more content, drafts, or code, you already know the feeling: the machine keeps up fine, but now you're the one struggling to keep quality high and direction clear.

Claude Code sits in the same category as agent-driven coding tools that turn plain instructions into real changes across a codebase. Cowork extends that idea toward broader work. The distinction from a plain chat interface, or from calling the OpenAI API in a script, is that these tools are built to take action across many steps rather than answer one question at a time. That's exactly why the review and coordination problem gets bigger as they get better.

The skeptical note worth holding onto: an 8x code figure comes from the company that makes the tool, describing its own teams. Treat it as a directional signal about how fast execution can scale, not a benchmark you should expect to hit. The more durable insights are the structural ones, that planning matters more, review becomes the constraint, and the collaborative texture of engineering work changes in ways that aren't all positive.

If you want help figuring out where agent-driven coding or AI workflows fit in your own team, browse the provider directory to find people who can put these ideas to work.