Fiona Fung | Lenny's Summary

1. Key Themes

Coding Is No Longer the Bottleneck — Verification Is the New Constraint

The Claude Code team has seen 8x more code shipped per engineer quarter-over-quarter, but the constraint has shifted from writing code to verifying it. With more disciplines — PMs, designers, data scientists — now committing code, the quality and verification problem is the next frontier.

"Coding is no longer the bottleneck. It's lifted the ceiling of what anyone is able to do." 00:00:05

"When not only more people checking in code, but like kind of different disciplines, but also the throughput is so high. How do we think about verification? That's kind of this other shift that I'm seeing." 00:09:06

The Shift from Synchronous to Asynchronous Engineering Work

The future of engineering is agents running in parallel on your behalf, not you actively prompting. Fiona describes how "routines" allow her to spawn agents overnight that generate PRs and summaries for morning review — a fundamentally different relationship with work than flow-state coding.

"We're shifting more towards async, like asynchronous. And so to your point, the fleets of agents, that's why routines are so interesting. Cause it's almost like I used to be doing a prompt synchronously, and then I would maybe kick off different prompts asynchronously, but now I can actually have a routine that actually generates these prompts for me." 00:35:35

"I wake up and I ended up having PRs that I could review versus before... it's like that higher level abstraction of, okay, now how can I actually write a routine that also basically does prompts for me for spawning different agents." 00:36:50

Specs and Frameworks Checked Into Repos Are the New TDD

Rather than relying on human reviewers for quality, Fiona's approach is to encode "what good looks like" directly into the repository as specs, then have Claude Code review against those specs automatically. This is described as an evolution of test-driven development.

"Claude is very good when you give it a framework to validate against those frameworks... if you have specs, check those into the repo and then make sure the spec also keeps up to date with the code frequently. But that's what I found works really well — any time you have a statement of what good looks like, get it into the repo and then Claude Code review can make sure it's still matching what you set up to do." 00:15:18

"For TDD, the test first principle is really good. But I think I myself struggled a bit because it was almost like you have to eat the broccoli first... the fact that test generation used to be this tax that I remember having to pay — that's not automated and you can even revisit all these principles that have been around for a while, but now they actually might be even more efficient." 00:15:56

High Agency Must Be Paired With High Accountability

The Claude Code and Cowork team's operating philosophy is that the freedom to execute independently is only valuable when matched with ownership of outcomes. This is not just a cultural value but a structural expectation.

"We really, it's about like, hey, here's a problem. And it's really everybody on the team has ideas for how to address the problem. So it's really high agency. And then we say with high agency is also high accountability. So it's all about making sure folks have that freedom to cook. But then it's also like, okay, what's the accountability for it as well." 00:38:47

The Two Hiring Profiles That Actually Work in an AI-Native Team

Rather than generalist engineers, Fiona has converged on two specific archetypes: creative product builders with strong product sense, and deep systems experts who can verify what AI generates. Both are essential and neither is interchangeable.

"The deep subject matter expertise — for example, when I first joined Claude Code, we had really great kind of like product generalists. And then I realized, oh, we were missing folks with a systems background. And so that was definitely an area that we needed more folks with kind of like systems and distributed systems expertise." 00:17:18

"The product sense folks, almost like the dreamers. These are folks that usually will be like, my gosh, you're really passionate about a product and they have an idea, they build it. And then it's always looking at the feedback and then iterating and polishing and making sure that the product is a delightful experience." 00:17:44

Dog Fooding Is the Highest-Signal Source of Product Intelligence

Across every product Fiona has run — Visual Studio, Facebook Marketplace, Meta VR, Claude Code — using the product herself has surfaced insights that neither metrics nor dashboards could. She treats dog fooding as a core leadership practice, not a nice-to-have.

"Every time I watch someone work, I learned something myself as well." 00:57:49

"After I left the team, one time I had a MacBook Air I wanted to sell. I could not believe it — the minute I put it up for sale, a buyer tried to scam me. And it was an interesting new scam vector I didn't detect. And so that goes to, again, like people will use your products in ways that you may not expect." 01:04:36

The Loneliness Problem: AI-First Teams Need Deliberate Human Connection

As engineers spend more time working solo with agents, the social fabric of engineering teams has started to fray. Fiona identified this early and introduced structured pairwise programming lunches as a direct response.

"After a while we felt it could start being a lonely experience because we all started just working with our agents so much. So recently we started a pairwise programming lunch. And because what we also learned was on Claude Code, everybody uses Claude Code and Cowork in such a different flow. And so we found that, wow, when we do pairwise programming, we actually learn so much from each other." 00:56:34

Latent Demand as a Product Discovery Framework

Anthropic consistently identifies massive opportunities by watching for behavior that wasn't intended — non-coders using Claude Code led to Cowork; small business owners using Cowork for invoicing led to Claude for Small Business. The pattern is: watch for people jumping through hoops, then smooth the path.

"With coding as a use case, so many of us were our own first customers. But latent demand has been like a framework — for example, with Cowork, we noticed a lot of folks that were not necessarily coders were using Claude Code. Can we make that experience better?" 00:32:42

"When you see people jumping through hoops to make something work, can you actually make that an even smoother and better experience?" 00:34:57

Explicit Permission to Kill Processes That No Longer Serve the Team

One of the explicit cultural pillars of the Claude Code and Cowork team is giving people formal license to abandon workflows and processes that have become obsolete. In a landscape changing this fast, holding on to old processes is actively harmful.

"One other thing that's really big on Claude Code and Cowork team culture is explicit permission to kill processes that no longer serve us." 01:23:01

2. Contrarian Perspectives

The Metric You're Measuring May Be Actively Misleading You

Most engineering leaders treat productivity metrics as ground truth. Fiona argues the opposite: the faster things change, the more likely your current metric is measuring the wrong thing. She uses the Facebook Marketplace "number of sellers" example to show how a metric that made sense at launch actively blocked the right expansion decision.

"Don't forsake motion for progress. Because if you're measuring tool usage, then you're measuring the action, but is it really making whatever the end outcome of yours important? Whatever metric, whether for productivity or even for product, always keep an eye and make sure that you're not just having blinders on blindly following a metric that used to make sense. Cause sometimes the landscape can change so fast, even the metrics themselves might need to be adjusted." 00:41:21

"In this area, the number of sellers is low, but actually people are finding items that they're looking for, which is what we're aiming for. And then I realized in that region, it wasn't a large number of sellers, but there were power sellers. But our first gate before we expand would have just been like, you know, factoring heavily number of sellers." 00:42:42

Managers Who Start as ICs First Build Stronger Teams

Conventional wisdom says experienced managers should immediately step into the management role. Fiona's practice is the opposite: require new managers to spend time as individual contributors first, without people-management responsibilities, so they earn credibility through code and product understanding before earning authority over people.

"Before you have to take on that full responsibility, give yourself that maker time to actually tap deep into the code and learn the code base. It's important because it keeps me in the flow. As amazing as metrics and everything are, if you as a leader are not living and breathing your product every day, you sometimes kind of lose touch of the touch and feel of the product." 00:50:52

The Engineers Resisting AI Are Drawing the Wrong Lesson From Early Model Failures

Engineers who tried AI tools at Sonnet 3.5 and dismissed them based on errors are now stuck with an outdated prior. The rate of model improvement means past failures are poor predictors of current capability — and this dynamic will continue.

"The one thing was, I remember the first time, it was probably Sonnet 3.5 or 3.6, and I remembered it was still making some mistakes when I was doing things on the side. And then I noticed some of the engineers that were resisting AI tooling, they're like, but see, look at all of these. But then I think it was hard to understand the exponential rate of improvements. Always thinking about what may have not worked — it might be worth the time to revisit because now might be a new capability." 01:17:27

The Next Generation of Engineers May Need Apprenticeships, Not CS Degrees

The traditional on-ramp to engineering — years of writing code, debugging, learning fundamentals through repetition — may be broken by AI. Fiona suggests the field may need to move toward fellowship or apprenticeship models, but is openly uncertain about what replaces foundational skills you never had to learn consciously.

"I wonder if it's for software engineering, it's almost like you go more towards a fellowship or apprenticeship program. I know it's odd because technically we have internships that would, but those were like three months and little projects. But I do wonder, and I wish I had a crystal ball here, how do you almost cram in some of the life experiences that we all got? How do you actually enable us to teach that to the next generation of builders?" 01:13:49

Anecdotes Beat Data When You Have Both

Counterintuitively for a data-driven engineering leader, Fiona consistently finds that firsthand product experiences reveal more than dashboards — and credits this as a leadership superpower.

"Sometimes it's also how I've been able to most effectively help the team. I wanted to use my dog fooding time to actually vet how the experience looks. That was a way that I felt I could still meaningfully contribute to help hold the quality bar for the team." 01:06:02

Lenny adds: "I think it was Jeff Bezos that said, if you have the data and you have an anecdote, trust the anecdote over the data, surprisingly." 01:08:24

3. Companies Identified

Anthropic AI safety company, maker of Claude. Discussed as the primary subject throughout — specifically the Claude Code and Cowork teams, their engineering culture, and the 8x code output increase.

"Anthropic engineers on average ship eight times as much code per quarter as they did compared to 2025." 00:03:10

WorkOS Enterprise B2B SaaS infrastructure company providing SSO, SCIM, RBAC, audit logs.

"If you're building a product for the enterprise, you've felt the pain of integrating single sign-on, SCIM, RBAC, audit logs, and other features required by large companies. WorkOS turns those deal blockers into drop-in APIs with a modern developer platform built specifically for B2B SaaS. It's essentially Stripe for enterprise features." 00:06:49

Mercury Banking platform for entrepreneurs and startups, recently launched Command — a conversational financial interface.

"Does your bank have an API, a terminal native CLI, or an AI ready MCP server? I don't think so. And just recently they launched Command, a conversational interface built directly into Mercury, which acts as your financial operator." 00:48:45

Facebook / Meta Discussed as the company where Fiona built Facebook Marketplace from zero to over $100 billion in GMV annually, and led VR/AR hardware product teams.

"Today, Facebook Marketplace generates over $100 billion in GMV every year." 00:01:04

Microsoft / Visual Studio Discussed as the foundational context for Fiona's early career on the TypeScript and Visual Studio teams, and the origin of her dog fooding philosophy.

"I use a VS editor to build the VS editor. And that's where my whole love of dog fooding comes from." 00:05:48

IBM Discussed as Fiona's origin as an engineer working on DB2 operating system services, pre-IDE, working in Vim on terminal.

"IBM working on DB2, the operating system services team. Back then I was thinking, oh, how can I be like the... what's a hard area of the stack?" 00:04:03

Instagram Discussed as the company where Fiona led infrastructure, growth, integrity, and safety teams, overseeing an org of over 500 people.

"She went to Instagram, where she led infrastructure growth, integrity, and safety teams. While at Instagram and at Meta, she oversaw an org of over 500 people." 00:01:32

Cursor Mentioned in passing as a WorkOS customer and example of a winning AI-native company.

"What do OpenAI, Anthropic, Cursor, Vercel, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS." 00:06:49

Replit Mentioned as a WorkOS customer.

"What do OpenAI, Anthropic, Cursor, Vercel, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS." 00:06:49

Vercel Mentioned as a WorkOS customer.

"What do OpenAI, Anthropic, Cursor, Vercel, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS." 00:06:49

OpenAI Mentioned as a WorkOS customer and context for competitive AI lab landscape.

"What do OpenAI, Anthropic, Cursor, Vercel, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS." 00:06:49

Sierra Mentioned as a WorkOS customer.

"What do OpenAI, Anthropic, Cursor, Vercel, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS." 00:06:49

Clay Mentioned as a WorkOS customer.

"What do OpenAI, Anthropic, Cursor, Vercel, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS." 00:06:49

4. People Identified

Boris Cherney Head of Claude Code at Anthropic. Credited with hand-rolling the original Claude Code codebase and with the much-cited line "coding is solved."

"Boris famously said, coding is solved." 00:03:10 "You have Boris who hand-rolled the code in the early days. And of course now does anymore, but he gained that knowledge from before because he was in the code days." 00:55:36

Kat Wu Leader of the Cowork team at Anthropic. Mentioned as one of the top-listened guests on Lenny's Podcast.

"She oversees both Boris Cherney and Kat Wu, both of whom who have been on the podcast and whose episodes are in the top 10 most listened to episodes of all time." 00:01:04

Mohamed Hegazi Anthropic team member who contributed topic suggestions for this episode.

"A huge thank you to Kat Wu, Boris Cherney, and Mohamed Hegazi for suggesting topics and questions for this conversation." 00:01:59

Tyler Cowen Economist and podcaster. Referenced for his concept of "initiative" as the defining trait of people thriving in the AI era.

"Tyler Cowen had this awesome talk about just like what's happening. The people that are doing best in this world — his term is initiative. They have initiative." 00:38:01

Sheryl Sandberg Former COO of Meta. Cited for her advice on culture during hypergrowth: rapid scaling is a problem you want to have.

"Her advice was, this is actually the problem you want to have, because this means you're growing and doing well. And this is normal versus you can, nothing will change if you're doing badly. That's a much worse situation." 01:20:51

Jeff Bezos Founder of Amazon. Cited for his counterintuitive prioritization of anecdotes over data when both are available.

"I think it was Jeff Bezos that said, if you have the data and you have an anecdote, trust the anecdote over the data, surprisingly." 01:08:24

5. Operating Insights

Use Claude as a Management Intelligence Layer, Not Just a Coding Tool

Fiona has set up a persistent Claude Code remote session enrolled in all repos, with access to Slack and metrics dashboards. She uses it not to generate PRs but to prepare for 1:1s — reviewing what shipped, what broke, what the impact was, and identifying hotspots before the conversation.

"I have a Claude Code remote session that I enlist in all of our repos. This way, I have full visibility into the work that everybody's doing. This instance also has access to all our Slack channels and I'll have access to how are the metrics of everything we track. And so every month, we'll actually do it together, share my screen and do our Claude Code session. And it's just about, hey, what were the focus areas? What were some of the products that got shipped? How did it do? And so I have these sessions to enable me to have conversations with folks that I support." 00:10:05

The Bad/Sad Quality Framework: Classify Failures Before Measuring Them

Rather than just tracking raw performance metrics, Fiona introduced a two-tier classification of user experience failures: "bad" (irrecoverable, e.g., crash, data loss) and "sad" (recoverable pain points, e.g., flickering). Each team owns defining what counts as bad and sad for their surface, enabling cross-team quality rollups without dashboard noise.

"I started this concept of what's bad versus what's sad. And bad is like a very bad irrecoverable error. And sad is something that's kind of like a pain point recoverable. But when you stack up sads, it could generally go to bad. Even having a high level framework like that — versus just raw performance or reliability numbers — having some framework of what we think is a bad experience and making sure we're focused on addressing those." 00:45:16

Build Routines That Prompt the Prompts — The Morning Ritual Upgrade

Fiona replaced her manual morning Slack scan with an automated routine that runs overnight, synthesizes feedback channels, identifies themes, and generates draft PRs for polish fixes — so she arrives at work with PRs to review, not a raw inbox to parse.

"I have a routine that I set to run every morning at a certain time. It's able to actually go and kick off agents on your behalf. Hey, look at these feedback channels. And then if you're seeing some of these bugs, what are some polish fixes that you might be able to knock out? And then it would go kick off, and then I wake up and I ended up having PRs that I could review." 00:36:21

Pairwise Programming Lunches to Combat AI-Era Loneliness and Share Tacit Knowledge

As solo agent-based work replaced collaborative coding, the Claude Code team found that individual engineers had developed radically different personal workflows — and that watching each other work was unexpectedly high-value learning. Structured pairwise programming sessions became the vehicle for both reconnection and workflow knowledge transfer.

"After a while we felt it could start being a lonely experience because we all started just working with our agents so much. So recently we started a pairwise programming lunch. And because what we also learned was on Claude Code, everybody uses Claude Code and Cowork in such a different flow. And so we found that, wow, when we do pairwise programming, we actually learn so much from each other." 00:56:34

Give Explicit Permission to Kill Processes — Make It a Cultural Norm

In fast-changing environments, teams default to maintaining existing processes out of habit or social friction. Fiona makes discontinuation an explicit, named, cultural right on her team so people don't feel the need to justify stopping things that no longer serve.

"One other thing that's really big on Claude Code and Cowork team culture is explicit permission to kill processes that no longer serve us. And so maybe a suggestion is for anyone, any team — that suggestion would be great." 01:23:01

6. Overlooked Insights

Claude for Small Business: A Product Line Born From a Single Manager's Personal Experiment

Fiona casually mentions that after personally helping a few small business owner friends use Cowork for invoicing, expenses, and menu pricing, Anthropic subsequently launched "Claude for Small Business" as a product. This is a product discovery pattern worth noting: a senior leader's personal use case, validated informally with real users outside the company, became a product line. The mention is brief and throwaway, but it reveals that some of Anthropic's market expansion decisions are driven bottom-up from informal dog fooding by leaders, not formal market research.

"Actually, it's funny — you mentioned like the Claude for, uh, my passion with small business. After I had a few of those visits, we ended up now launching Claude for Small Business, which is really cool. And it totally didn't come from me, so I'm not taking credit, but I noticed it too — when I was working with them, because they were asking me, oh, does it have this plugin or this plugin?" 00:33:12

This matters because it suggests Claude for Small Business — a potentially enormous market (hundreds of millions of SMBs globally) — may have been underweighted as a strategic priority until an internal leader's personal relationships with small business owners created the signal. Investors and operators should watch this product line closely.

The Context-Switching Tax of Async Agents Is the Next Unsolved Productivity Problem

In a single brief exchange, Fiona identifies that the shift to async agent work has created a new cognitive burden she hasn't solved: she now has to block focus time to catch up on the async work she kicked off, reversing the original time-saving logic. This is a nascent but structurally important problem — as agent fleets scale, the review and context-restoration overhead could become a new bottleneck that replaces the old coding bottleneck. Nobody in the conversation digs into it, and Fiona explicitly says she hasn't cracked it.

"I used to book out like focus time for coding, cause you want that dedicated and then the whole context switching. And then it's interesting now that because I can context switch more with more async agents, I'm noticing I do actually have to go back and block like a focus time for me to catch up on all the different async work that I've kicked off." 01:12:12

"I haven't cracked it yet." 01:12:44

This is a product opportunity: the interface layer that helps knowledge workers triage, prioritize, and re-enter async agent work sessions without full context reconstruction is not yet built. The team that solves "async agent inbox management" could be the next major developer productivity company.