When Coding Stops Being the Bottleneck: Inside Anthropic's Claude Code and Cowork Teams

Fiona Fung has been an engineer for more than 25 years, starting with an internship at IBM, where she worked on DB2 using Vim and terminal debugging. When she joined Microsoft she landed on the Visual Studio team without knowing what an IDE was, and she describes seeing debuggers, breakpoints and multithreaded debugging as a mind-blowing step change. She spent about 11 years on Visual Studio and learned dogfooding by using the editor to build the editor, getting fast feedback from teammates who were heavy users.

She recalls that software then shipped on CDs, with hard manufacturing deadlines that made engineering time precious and planning heavy. That constraint has now flipped. Fung says that on the Claude Code and Cowork teams, coding is no longer the bottleneck, and that designers, PMs and everyone else on the team now check in code.

With more people checking in code and far higher throughput, she says the key question becomes how to verify it all. Lenny Rachitsky cites the tweet showing Anthropic engineers shipping about eight times as much code per quarter as before. Fung treats verification as the shift to watch.

Fung says she runs a Claude Code remote session across all of the team's repos, with access to the Slack channels and the metrics the team tracks. Each month she and her reports review it together: what the focus areas were, what shipped, how it did in market, and what the feedback channels said. She uses this to start conversations rather than to generate pull requests.

Her motto is "make new mistakes," because aiming for zero mistakes probably means moving too slowly. She also asks Claude to look across incidents for themes and hot spots, which she says used to be a manual process she could not have done a year earlier. Rachitsky notes that this is one of the few ways to keep up with eight times the output.

Her morning ritual used to be reading feedback channels over coffee, which collect emails, social posts and internal comments. A month or two before the conversation, Anthropic launched routines, and now she has one that summarizes feedback overnight and drafts pull requests for her to review. She says it is like having an agent write the prompts and the PRs.

Human review had become a major bottleneck, and Claude code review did not even exist the year before. Fung says areas that need deep subject-matter expertise still get human reviewers, but Claude is good at validating against a framework. Her advice is to check a statement of what good looks like into the repo, keep it current, and let code review confirm the code still matches it.

Rachitsky calls this the evolution of test-driven development. Fung agrees, recalling that writing tests first felt like eating broccoli, whereas now Claude writes the failing test and then the fix. She says the old test-writing tax is now automated, so older principles may be even more efficient now.

On hiring, she looks for creative builders with product sense and deep systems experts. When she first joined Claude Code the team had good product generalists but lacked distributed-systems expertise, and she says it is "trust but verify" wherever the models are not yet reliable. The builders are the dreamers who own a product end-to-end, iterate on feedback and polish it.

Both agree the common thread is ambition. Fung describes an engineer who is not a mobile specialist but took on a mobile version of a feature with Claude as a partner. In her words, AI has lifted the ceiling of what anyone is able to do.

Asked what separates people thriving with AI from those frustrated by it, Fung points to a growth mindset: always learning, even though what made you successful may not keep serving you. She first understood this fully in her move from Microsoft to Meta. She also notes that frustration is sometimes fear in disguise.

Her advice for fear is to lean in and ask what is within your control, so that events happen "for you" rather than "to you." She illustrates this with her teenage self, who wanted to study engineering but feared she could not afford it. She took a bank teller job posted at her high school, worked weekends, paid for her tuition and kept the job for two years during the dot-com downturn.

She also likes the questions "what would you do if you weren't afraid?" and "do something scary once in a while." When you are at peak efficiency, she says, doing something unfamiliar causes a dip but is how you keep growing. Rachitsky adds his most-used quote, that the cave you fear contains the treasure you seek.

Fung worries about a gap between people who lean into AI and those who do not. Her passion for small business traces to her grandmother, who found community at a Cantonese-speaking yarn shop after they emigrated from Hong Kong to Canada. She began using Cowork for her own travel expenses and then helped friends who run small businesses and spend their evenings on invoices and bills.

Watching them go through onboarding turned up real bugs, and they used the tool in ways she had not imagined. One friend who runs two restaurants asked Claude to find menus in a cluttered folder and then to compare prices with similar restaurants in the area. Her advice to listeners is to start with a meaningful way AI has helped them and use that as a conversation starter with friends, family or a favorite local business.

Asked how Anthropic spots big opportunities early, Fung says she cannot compare with other labs but her teams watch for latent demand. They noticed non-coders using Claude Code, which led to Cowork, and small-business visits led to a Claude for small business bundle inside Cowork. She says that when people jump through hoops to make something work, you can make that experience smoother.

Fung says the next frontier is asynchronous work. Where she once ran a prompt synchronously and later kicked off several prompts in parallel, a routine now runs on a schedule, checks things such as feedback channels, and spawns agents to prepare polish fixes. She wakes up to pull requests to review, and describes the level of abstraction as constantly rising.

Rachitsky summarizes this as writing down the daily tasks a manager does and having Claude do them and report back. Fung adds that with good verification you can give the agent even more autonomy and say "go for it."

On the initiative that economist Tyler Cowen says distinguishes top performers, Fung says agency is something the Claude Code and Cowork teams value highly. Her team's saying is that high agency comes with high accountability: people have freedom to cook, but should be able to state the hypothesis they are trying to prove.

On measuring the ROI of AI spend, Fung recalls the old debates over lines of code, throughput and time to land a PR, each of which had flaws. She sees token maxing as a version of lines of code and quotes a saying she likes: don't forsake motion for progress. She prefers to zoom out to the problem being solved and a measure of that.

Beyond metrics, she advises leaders to run a listening tour, especially among senior engineers, about what is working and what is not. She also says to check that a metric still serves its goal, using an early Facebook Marketplace example. Their gate counted number of sellers, but one region had few sellers who were power sellers and buyers were finding what they wanted, so they changed the metric.

For quality at high velocity she wants more proactive detection. She introduced the idea of bad versus sad: bad is an irrecoverable error and sad is a recoverable pain point, though stacked sads can become bad. Each team defines these for its own surface, for example a crash on the CLI is bad and flickering is sad.

Rachitsky concludes that monitoring and tests are the best way to stay on top of quality, which also lets agents close the loop themselves. Fung agrees and mentions the team's dashboard that tracks swear words as a signal of user frustration. She admits evals are hard because they have to capture the user experience as well as accuracy.

Fung has every new manager on her teams begin as an IC. The idea came from a listening tour in which people asked for clearer prioritization and fewer review layers. She says that before taking on the heavy responsibility of supporting people, managers should spend maker time going deep into the code or product, which builds rapport instead of reaching for the manager toolbox.

She does PRs herself, less for what she fixes than to use the product daily and keep her feel for it, because dashboards are not enough. Her teams often tell her they appreciate a leader who uses what they build. She did the same at Meta, where she was an IC for her first quarter to learn what it is like to ship there.

She says she had not shipped production code since about 2017 and worried about causing bugs. At Claude she used Claude as an onboarding buddy to explore the codebase, write automated tests and plan manual tests. Friends who have managed for years have told her they are shipping code again thanks to Claude.

On whether coding skills atrophy, Fung says the team debates it. Boris Cherny hand-wrote early Claude Code and gained knowledge that way, so she tells new engineers to keep understanding the architecture of their changes. Her principle is to keep "double clicking" into the layer you depend on, because trust but verify still applies.

The team also found that working with agents can become lonely, so they started pairwise programming lunches and hackathons. Everyone uses Claude Code and Cowork very differently, and watching teammates teaches her something new each time. She adds that the old joy of flow and the aha moment at the end of a gnarly problem is partly lost, and some engineers say the hardest parts are what they used to enjoy.

Fung says all coding-adjacent roles are shifting. PMs are no longer blocked on engineering bandwidth and have helped ship features, and design and data science are likely next. Engineers are becoming more product-minded, as the roles blur and fewer cross-functional blocks remain, which makes verification matter even more.

Fung says her obsession with using the product keeps her close to the dream the team is trying to enable. She gives examples from Meta: she put a MacBook Air on Facebook Marketplace and ran into a new scam vector, and while supporting VR she repeatedly found floor-height issues because of how she set up her space. People use products in ways you do not expect, she says.

Dogfooding also gave her a way to contribute meaningfully when she did not check in code: on a VR team she vetted the polish fixes and helped hold the quality bar. If you cannot use the product, she says, meet customers. On a research trip to Chile for Marketplace she found that slow LTE connections on Android phones prevented the feed from loading, which was the real growth blocker.

Rachitsky revisits three questions from Fung's talk: whether separate iOS and Android orgs are still needed, how far to push automated reviews, and how to keep everyone equally productive with blurring roles. She says the team still needs mobile experts, but not as large an org since people can flex. Verification of the experience, as opposed to accuracy, is the hard part to automate.

A new problem is context switching. With routines and async agents, the load on context switching is rising, and she now has to block focus time just to catch up on the work she kicked off. Rachitsky notes that agents can help you reset, so it is both better and worse. Fung says she has not cracked it.

On jobs, Fung says her big open question is how the next generation of engineers learns, since her path and Rachitsky's were so different. She wonders whether apprenticeship or fellowship programs could help people understand the layer beneath. She also suggests that what matters may shift over time, and advises revisiting things the models could not do before, since the next model may handle them.

What keeps Fung up at night is team culture, which she calls a living thing that shows up in how people treat each other. She wants diverse perspectives, open debate and a one-team mentality, meaning that near the finish line you look behind you to see who needs help. Rachitsky shares what worked at Airbnb, where founders constantly reinforced culture, and Sheryl Sandberg's advice that rapid growth is a good problem to have.

She asks managers to be open about what is not going well. Her nightmare is a manager who says everything is fine and then turns out to be struggling, like the meme of a dog drinking coffee in a burning room. She also describes a team norm of explicit permission to kill processes that no longer serve the team.

Her example is planning. She started with a lightweight six-month roadmap, but three months in found nobody was referring to it. Now the team does what she calls just-in-time planning: a monthly priorities spreadsheet with no docs, checked briefly each week, and the whole team gathers every six months to set themes. She is already thinking about automating even that so it never feels like a tax.