Fiona Fung says coding is no longer the bottleneck at Anthropic, where engineers produce eight times as much code per quarter as in 2025.
Fung says the new challenge is verifying that high volume of code, using a trust but verify approach to AI.
Fung says her team sets bad and sad thresholds for each product surface and tracks swear words as a frustration signal, because output should serve outcomes rather than lines of code.
When AI writes the code, judgment and verification become the job
Once AI agents write most of the code, the scarce skills at Anthropic become judgment, verification and knowing what to build. Fiona Fung describes how the Claude Code and Cowork teams adapt their management, hiring and metrics to that shift. Her view is that output only counts if it serves outcomes, and that people still need accountability, closeness to customers and each other.
Takeaways
Coding is no longer the bottleneck; verification is. Anthropic engineers produce about eight times as much code per quarter as in 2025. The hard problems at Claude Code and Cowork are now verifying that throughput, holding quality, and deciding what is worth building. Fiona's stance toward AI is trust but verify.
Managers use Claude to run their own work. Fiona runs a Claude Code remote session across all team repositories, with Slack and metrics access, to prepare monthly reviews. Routines, a scheduled feature, summarises feedback and prepares pull requests each morning, since she cannot track the feedback volume alone.
Explicit specs make Claude a strong reviewer. Claude Code reviews did not exist a year ago. Fiona says that giving Claude a clear framework or spec for what good looks like makes it a strong reviewer, which echoes the test-first thinking of the 2000s.
Measure outcomes, not lines of code. Fiona says output should serve outcomes and warns against lines of code or tool usage as metrics. Her team sets a bad (irrecoverable) and a sad (recoverable pain point) threshold for each surface, and tracks swear words as a frustration signal.
Hiring favors creative builders and deep specialists. Fiona looks for builders with product sense and for deep systems experts on hard problems. Anthropic still needs iOS and Android experts, though fewer per specialism. Product managers are no longer held back by engineering bandwidth.
High agency requires accountability and a hypothesis. Fiona says teams with high agency also carry high accountability and need a hypothesis about the problem they solve. Mistakes are fine if they are new ones, because aiming for zero mistakes means moving too cautiously.
Closeness to users and people still matters. Fiona relies on dogfooding and customer visits, which led to Claude for Small Business. Engineers working with agents felt lonely, so the team started pair-programming lunches and hackathons. How newcomers learn fundamentals remains open.
Advice for managers and individuals
Spend time as an individual contributor before you start supporting people.
Visit customers often, because every visit teaches something new.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Revisit ideas Claude could not do before, since a new model may now handle them.
Ask what is within your control, and do something scary once in a while.
Keep a culture with explicit permission to kill processes that no longer serve the team.