Skip to content
ClaudeClaude

Running an AI-native engineering org

When agentic coding goes from individual tool to org-wide default, the tool isn't the hard part…your processes are. Fiona Fung, Leader for Engineering and Product for Claude Code and Cowork, walks through how the bottlenecks changed at Anthropic (review, ownership, hiring) and the norms we rewrote to keep shipping.

May 22, 202626mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

How AI shifts engineering bottlenecks, norms, and org design practices

  1. Engineering bottlenecks have shifted from writing code to verification, review, and long-term maintenance as AI increases throughput.
  2. Claude Code’s team norms changed accordingly: fewer heavy design docs, more prototyping and PR-based debate, and a stronger emphasis on automated, “shift-left” quality checks.
  3. Code review becomes a hybrid system where AI handles style, lint, obvious bugs, and spec-drift checks while humans focus on legal risk, trust boundaries, and product taste.
  4. Org and hiring strategies adapt to blurred roles by prioritizing “creative builders with product sense” alongside deep systems experts, and by having managers start as ICs to dogfood and understand the codebase.
  5. Success signals include faster onboarding ramp-up, reduced burden on teammates for support, improved PR cycle time (diagnosed by funnel stage), and a high share of Claude-assisted commits—while still measuring real product outcomes beyond throughput.

IDEAS WORTH REMEMBERING

5 ideas

Audit norms continuously because AI changes what’s scarce.

Processes built to protect limited engineering time can “quietly stop working” when coding becomes cheap; regularly re-check whether each ritual still serves its original purpose.

Treat “code wins” as the new default for technical debates.

Instead of long whiteboard arguments, generate multiple PR/prototype variants quickly and debate concrete tradeoffs (implementation, team impact, user experience) using working code.

Reduce planning overhead, but increase verification rigor.

The team shifted away from extensive design docs toward prototypes and PR discussions, while doubling down on automation and shift-left testing to keep quality high with increased throughput.

Use AI to scale reviews, but keep humans for risk and taste.

Claude can catch style issues, lint, obvious bugs, and spec drift (especially if specs are checked into the repo), while humans remain essential for legal reviews, security/trust boundaries, and subjective product quality.

Make the codebase (and checked-in specs) the source of truth.

Documentation outside the update loop becomes stale faster when changes accelerate; keeping specs and key knowledge in-repo lets AI and humans reference an up-to-date canonical source.

WORDS WORTH SAVING

5 quotes

What served you prior may no longer...

Fiona Fung

Coding is rarely the slow part anymore.

Fiona Fung

In technical debate, code wins. You know, it's like building is cheap, arguing is expensive.

Fiona Fung

If Claude can do it, Claude should.

Fiona Fung

Pick your noisiest workflow... and always ask yourself, "Is this still serving its purpose?"

Fiona Fung

Bottleneck shift: coding to verification/review/maintenanceReplacing heavy planning with prototypes and PR discussionsAI-assisted code review vs human judgment boundariesDogfooding as a product-sense muscle (especially for managers)Role blurring and cross-functional gap-filling with AIOrg shape: flatter pods and managers as former ICsMetrics and proof: onboarding time, PR cycle funnel, CI capacity

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.