Anthropic Engineering Leader: ‘Coding Is No Longer the Bottleneck’ as Engineers Ship 8x More Code Per Quarter

Fiona Fung, who leads the Claude Code and Cowork teams at Anthropic, recently appeared on Lenny Rachitsky’s Podcast and offered an unusually candid look at what software engineering looks like inside one of the world’s most prominent AI labs. Her headline claim is direct: “Coding is no longer the bottleneck.” Fung backed that up by referencing an internal chart showing Anthropic engineers shipping 8 times as much code per quarter today compared with 2021 through 2025 levels.

Fung brings considerable credibility to that assertion. Before joining Anthropic, she spent 11 years at Microsoft building Visual Studio and TypeScript, then moved to Meta, where she helped start Facebook Marketplace (now generating over $100 billion in gross merchandise value annually), worked on Meta’s first VR and AR glasses, and led infrastructure, growth, and safety teams at Instagram. She has been a practicing engineer for over 25 years.

The New Constraint Is Verification

For two decades, software engineering productivity was measured in proxies: pull requests merged, lines shipped, and cycle time. Fung’s argument is that the cost of producing code has collapsed inside Anthropic, and the real constraint has moved downstream. As she put it, “Now it’s all about, where has that shift happened? Not only are more people checking in code, but like different disciplines, but also the throughput is so high, how do we think about verification?”

When a team can generate a quarter’s worth of code in a week, the scarce resource becomes the human and automated judgment needed to confirm that output behaves correctly, secures customer data, and matches intent. Code review, test infrastructure, observability, and post-deployment monitoring become the gating activities. Anthropic has already moved to address this directly: in March 2026, the company launched Code Review for Claude Code, a multi-agent automated PR review system that dispatches specialized AI agents to check for logic errors, API misuse, authentication flaws, and security vulnerabilities before code is merged.

Fung’s team also developed a practical monitoring shorthand called the “bad vs. sad” framework. “Bad” events are unrecoverable errors such as crashes. “Sad” events are recoverable friction points like UI flickering or a drop in conversation quality. Each team tracks both categories for its own surface areas, giving them a fast signal for identifying issues without waiting for formal evaluation cycles.

AI Is Blurring the Line Between Engineers and Everyone Else

Fung described a team composition that would have been unusual even a year ago. “We also have designers, PMs, everybody on the [Claude Code] team checks in code,” she said. If that description is accurate, it implies the traditional career ladder, where coding is a gated specialist activity, is being rewritten inside AI-native firms. Product managers prototype features. Designers ship UI changes directly. The engineering function shifts toward architecture, review, and systems thinking rather than raw production.

Lenny Rachitsky framed the broader trajectory for listeners: “People may forget 100% of code was written by humans not long ago. And now it’s getting to 100% of code written by AI.” That is clearly an overstatement, but it captures the direction Anthropic is experiencing internally.

There is also an unexpected social cost. Fung acknowledged on the podcast that as engineers increasingly work independently with their own teams of agents, isolation has emerged as a genuine challenge. “After a while, we felt it could start being a lonely experience because we all started just working with our agents so much,” she said. Anthropic responded by introducing hackathons to keep the team interacting, along with pairwise programming lunches where engineers work side by side, not necessarily on the same project, and pick up new patterns by observing how colleagues use Claude Code.

What an AI-Native Team Looks Like

Fung shared a practical example of how she manages her team day to day. She maintains a persistent Claude Code session connected across repositories, Slack channels, and internal metrics. On a monthly cadence, she said, “Every month… we’ll actually do it together. I’ll share my screen, then we do our Claude Code session” to review what shipped, how it performed, and what the feedback looked like.

The workflow resembles having a permanently available engineering analyst embedded inside the organization. Rather than spending time gathering information, managers focus on interpreting results and deciding what to build next. The bottleneck is no longer writing code; it is asking the right questions about what the code actually does.

What Investors Should Watch

Anthropic remains private, so its shares are not directly accessible to most investors. The more consequential takeaway is what Anthropic’s experience suggests about the broader AI ecosystem and where value accrues as code generation becomes abundant.

If verification and observability are the new scarce resources, the biggest beneficiaries may be companies building developer security, testing infrastructure, and evaluation tooling. The hyperscalers running the underlying AI compute also stand to gain as software teams push ever more workloads through AI models.

Anthropic is already cited as a strategic partner of Broadcom (NASDAQ:AVGO | AVGO Price Prediction), alongside Google (NASDAQ:GOOGL) and OpenAI, in the AI semiconductor ecosystem. A ProShares “FAB 10” ETF, which filed a preliminary prospectus with the SEC in June 2026, aims to package private AI leaders like OpenAI and Anthropic alongside public names such as NVIDIA (NASDAQ:NVDA), Microsoft (NASDAQ:MSFT), and Tesla (NASDAQ:TSLA), though the fund remains in pre-effective registration as of this writing.

The central question Fung’s data raises is whether Anthropic’s 8x throughput gain is specific to an AI-native organization that builds the tools its own engineers use, or an early glimpse of where the broader software industry is heading. If it proves repeatable at scale, the implications extend well beyond coding assistants and into every corner of enterprise software.

Editor’s note: This article was updated to include Fiona Fung’s professional background at Microsoft and Meta, Anthropic’s “bad vs. sad” verification tracking framework, the loneliness challenge Anthropic identified among AI-heavy engineering teams and the interventions it introduced, Anthropic’s March 2026 launch of multi-agent Code Review for Claude Code, and the current pre-effective registration status of the ProShares FAB 10 ETF.

Contact [email protected] for any questions or corrections.