All podcasts / Lenny's Podcast / Summary

What happens after coding is solved? | Fiona Fung (Claude Code & Cowork)

2026-06-21 - 99 min - source - Read full transcript
Lenny Rachitsky (host)Fiona Fung

Key insights

Anthropic engineers now ship roughly 8x more code per quarter than in 2025, which has shifted the real bottleneck from writing code to verifying it.
Fiona cites an internal chart showing flat output for years then a sharp spike. She frames this as raising the ceiling of what any individual can attempt, not just speeding up existing work.
ai-native-engineering
When coding stops being the constraint, ambition becomes the differentiator.
She describes engineers who previously dismissed feature ideas as 'too hard' now saying 'I just ask Claude to do it' -- the limiting factor shifts from technical feasibility to how big someone is willing to think.
ai-native-engineering
Two hiring profiles now matter most on AI-native teams: creative builders with product sense, and deep systems experts for the parts that still need irreducible human judgment.
Cloud Code initially over-indexed on generalist builders; Fiona found she needed to deliberately add distributed-systems and infrastructure expertise for areas where models still need close verification.
engineering-management
Managers can replace manual status-tracking with an always-on Claude Code session that has repo and Slack access, turning check-ins into conversations about impact rather than raw output.
Fiona keeps a persistent Claude Code instance across all of the team's repos and Slack channels. Monthly, she reviews it with each report to discuss what shipped, how it performed, and what feedback came in, rather than just listing completed tickets.
engineering-management
Automated code review works best when it validates against an explicit written spec checked into the repo, not against vague notions of quality.
She points to content-design guidelines checked into the codebase as an example: Claude reliably catches regressions once there is a concrete 'what good looks like' artifact to check against, extending the discipline of test-driven development into review.
verification-at-scale
A simple two-tier severity framework, 'bad' versus 'sad,' scales quality triage better than raw performance dashboards across many product surfaces.
'Bad' means an irrecoverable error (e.g. a CLI crash); 'sad' means a recoverable pain point (e.g. UI flicker). Each team defines its own bad/sad list, and the team watches for sads compounding into bads over time.
verification-at-scale
Routines let a manager automate her own daily rituals into async agent jobs that surface a summary and ready-to-review PRs by morning.
Fiona used to manually scan feedback channels over coffee each morning; a routine now does that scan and drafts fixes overnight, moving her role from generating prompts herself to reviewing what an agent already proposed.
ai-native-engineering
New managers on the team start as ICs, without people-management responsibility, before taking on reports.
This gives them time to learn the codebase and product deeply and build rapport as a teammate first. Fiona says this consistently produces stronger trust than managers who lead with 'manager toolbox' behaviors from day one.
engineering-management
Working almost entirely through individual agent sessions made engineering lonelier, so the team deliberately rebuilt shared time.
They started a 'pairwise programming lunch' (parallel work, shared space, not literal pair programming) and kept hackathons, because everyone uses Claude Code and Cowork so differently that simply watching a teammate work teaches something new.
team-culture-and-flow
Long-range roadmaps stopped working, so planning shrank to JIT (just-in-time) monthly cycles re-verified weekly.
A six-month roadmap doc she introduced was already stale three months in because the landscape moves too fast. The team now keeps a lightweight monthly priority list on a shared spreadsheet, checked against reality every week, and she is still looking to automate even that.
engineering-management
The way to close the AI-adoption gap is a personal, one-on-one story about a concrete use case, not a broad campaign.
Fiona's own on-ramp for small-business owner friends was showing them Cowork solving something she personally disliked (expense reports). She argues 'AI-pilled' people should each find one person in their life and share a specific example that changed their own workflow, rather than trying to explain AI abstractly.
ai-adoption-divide
Anecdotal evidence caught product failures that aggregate metrics missed, reinforcing that leaders should trust direct observation over dashboards alone.
During Facebook Marketplace's Chile launch, a low 'number of sellers' metric looked like failure until an in-market visit revealed slow LTE was breaking the feed, and that the region actually had healthy power sellers -- the original metric was measuring the wrong thing.
verification-at-scale

Books referenced

Media referenced

Companies

Techniques and frameworks

Summary

Fiona Fung, who leads the Claude Code and Cowork teams at Anthropic (and previously ran Visual Studio's editor team at Microsoft and founded Facebook Marketplace at Meta), walks Lenny through what engineering management looks like once coding itself stops being the bottleneck. The starting data point is stark: Anthropic engineers now ship roughly 8x more code per quarter than in 2025. Her core argument is that this doesn't just speed up existing work, it raises the ceiling of what anyone can attempt, and the constraint shifts from "can we build it" to "how ambitious are we willing to be" and, immediately after that, "how do we verify all of this is actually good."

Much of the conversation is about the verification and management practices she has built to keep pace with that throughput. She runs a persistent Claude Code session wired into every repo and Slack channel, which she uses in monthly reviews with reports to talk about outcomes rather than just count shipped PRs. She has replaced manual morning feedback triage with "routines," scheduled async agents that summarize feedback and draft fixes overnight. For code quality specifically, she pushes teams to check explicit specs ("what good looks like") into the repo so automated review has something concrete to validate against, and the team uses a simple "bad versus sad" severity framework instead of raw dashboards to keep quality triage legible across very different product surfaces.

On the people side, Fiona describes deliberately reshaping who gets hired: fewer generalist product builders, more mix of "creative builders with product sense" and "deep systems experts" for the parts of the stack that still need hands-on human judgment. New managers on her team start as individual contributors, without direct reports, so they build codebase fluency and rapport before taking on people-management duties. She's candid that working almost entirely through solo agent sessions has made engineering lonelier, prompting the team to invent "pairwise programming lunches" and keep running hackathons specifically to rebuild shared context and connection.

Planning has compressed dramatically. A six-month roadmap she introduced was stale within three months, so the team moved to "JIT (just-in-time) monthly planning": a lightweight, weekly-rechecked spreadsheet of priorities rather than a fixed long-range plan. She frames this, and most of her management philosophy, around a repeated growth-mindset principle: constantly ask whether a given process still serves its purpose, because the ground truth changes faster than any static plan can track.

Fiona also spends real time on the AI adoption gap outside of Anthropic, drawing on her own experience helping small-business-owner friends use Cowork for tasks like expense reports and menu search. Her advice for closing that gap is intensely personal rather than top-down: find one concrete use case that changed your own life and share it one-on-one, because abstract pitches about "AI" don't land the way a specific story does. Throughout, she repeatedly favors anecdotal, on-the-ground evidence over dashboards alone, citing a Facebook Marketplace Chile launch where an in-person visit uncovered a load-time failure and a mis-specified success metric that aggregate data had missed entirely.

The episode closes with a lightning round covering her book and film recommendations (Margaret Atwood, Haruki Murakami, The Little Prince, Amelie, and two Studio Ghibli films including Nausicaa of the Valley of the Wind, whose heroine she says shaped her leadership style), and a story about knitting through meetings, a habit she picked up from her grandmother and still practices during high-stakes conversations at Anthropic.

Notable Quotes

"Coding is no longer the bottleneck. It's lifted the ceiling of what anyone is able to do." - Fiona Fung

"We say with high agency is also high accountability." - Fiona Fung

"Don't forsake motion for progress." - Fiona Fung

"In a world where you can be anything, be kind." - Fiona Fung