Head of Claude Code: What happens after coding is solved | Boris Cherny

I ship dozens of pull requests a day with multiple agents running, and I haven’t hand-edited a line since November because coding is finally free of the minutiae.

My guest is Boris Cherny, who leads Claude Code at Anthropic; it has upended software work in a year, so why did you briefly leave for Cursor and then come back?

Cursor’s team is great, but I missed Anthropic’s mission, and the shared focus on safety is what I need to feel fulfilled.

Adoption has been explosive, with eye-popping commit stats and senior engineers barely touching code; how do you reflect on this first year?

The growth curve keeps steepening, and what began as a terminal hack tied to our safety-first roadmap spread across platforms as latent demand and rapid user feedback shaped it.

We think in exponentials, and by simply extending the curve I expected one hundred percent AI-authored code for many engineers; that’s been my reality since November.

Breakthroughs come from room to play, not rigid roadmaps, so I prototyped obsessively until the pattern was undeniable.

I still review outputs, and Claude now auto-reviews every pull request, with humans keeping meaningful checkpoints.

What’s the next big shift beyond one hundred percent AI-written code?

Claude is starting to suggest what to build by reading feedback and telemetry, and agents are taking on non-coding tasks too, which makes coding feel close to solved.

I point it at our Slack feedback channel and it proposes fixes and opens pull requests, which lets us close loops in minutes.

Model gains have been dramatic, with transfer from non-code training helping; productivity per engineer is up around two hundred percent, which is unheard of compared to the tiny year-over-year lifts I saw in Big Tech.

To avoid old habits, we under-resource projects to force automation, ship faster by default, and give engineers generous tokens early, then optimize once something works.

I enjoy coding more than ever because code is a means to build, though some still hand-code for the craft and that’s great too.

Skills live on a shifting stack of layers, and while understanding the layer below still helps now, it may matter far less soon.

The printing press is my frame: democratization explodes output and frees people to focus on what to create, even as the transition is painful.

Which roles are next, and what does this mean for jobs?

Product, design, data science, and most computer-based work will change as agents act through tools, and while we’re hiring, we’re also leaning into safety and policy because the societal impact is huge.

You’ve mentioned latent demand and Co‑Work; tell that story.

People kept using Claude Code for non-coding work, so a small team built Co‑Work in about ten days with Claude Code end-to-end, added a secure VM and guardrails, released it as a research preview, and it clicked immediately.

Safety means aligning and lab-testing, then learning from real-world use; we invest in interpretability and open-source a sandbox to help raise the industry bar.

Day-to-day I keep several agents running across terminal, desktop, and phone, and a surprising chunk of my “coding” now happens on mobile.

Leave us with practical guidance for builders and users.

For builders: don’t over-orchestrate, give the model tools and a goal, bet on more general models, and design for capabilities six months out as tool use and unattended runtimes improve.

For users: pick the most capable model with maximum effort, start in plan mode before execution, and try different interfaces since it’s the same agent everywhere.

Competition is healthy, but my focus stays on solving users’ problems and iterating quickly.

One last one, straight from Ben Mann: what’s your plan after AGI? What does life look like then?

I’d slow way down and make miso. Living in rural Japan taught me patience and long horizons, and fermenting miso forces that mindset in a way coding never does.

I love that. Before our lightning round, any final thoughts you want to leave people with?

At Anthropic we’ve long believed in a path from coding to tool use to full computer use, because that’s where capability and safety learning compound. Quad code getting huge is surprising in form but not in spirit, and it still feels early—most people don’t use it yet.

Wild to see the numbers and still feel like day one.

The growth comes from users who love it and push us with feedback. My favorite days are talking to them, fixing what’s broken fast, and, yes, waiting on a batch of miso.

Lightning round. What books do you recommend most?

Functional Programming in Scala for elegant thinking in types, Accelerando for a rocket‑paced take on our moment, and The Wandering Earth short stories for a fresh, beautifully written sci‑fi perspective.

Sci‑fi really trains how we picture the future.

Living by the seasons in Japan while reading sci‑fi made the stakes feel real, and that pulled me to Anthropic to help steer things a bit better.

Recent movie or show you liked?

I rarely watch TV, but I really enjoyed the Three‑Body Problem series.

Favorite recent product?

The Acquired podcast is fantastic for business history. And on our side, co‑work now really clicks—it reminds me of quad code’s early days, but it’s breaking through faster.

There’s also a Chrome extension so you can just talk to Claud in the browser.

Try co‑work by installing the quad desktop app and opening the co‑work tab. Start with a simple tool task like cleaning your desktop or drafting replies to your top emails, then connect tools so it can move info between mail, sheets, and Slack, and let it run tasks in parallel while you step away. For example, it pings teammates who miss weekly status updates so I never have to chase them.

Real‑world examples help people see what’s possible.

Your post with fifty non‑technical quad use cases became a co‑work benchmark; once it handled forty‑eight, we knew it was ready.

Go‑to life motto?

Use common sense. Don’t follow process blindly; reason from first principles and walk away when something smells off.

Why’d you get so active on Twitter, and how’s it been?

I’d been on Threads, but during a coding vacation in Europe I opened Twitter out of boredom, started replying to quad code posts, and turned it into a live bug and feature loop. With quad, I can often ship a fix in minutes, which surprises people, and the back‑and‑forth has been great.

Where can folks find you and be helpful?

Find me on Threads or Twitter.

Thanks for listening. Subscribe on your favorite app, leave a rating or review, and find past episodes at Lenyspodcast.com. See you next time.