Anthropic Publishes Playbook for Running an AI-Native Engineering Org

Most discussions of AI-assisted coding focus on the tool. Fiona Fung, Director of Engineering for Claude Code at Anthropic, is more interested in what happens to your organization when the tool works.

Her talk at Code w/ Claude SF 2026 — now published as a written post on Anthropic’s blog — is one of the most honest accounts yet of what it actually takes to restructure an engineering team around agentic AI. Not what the marketing deck says. What breaks, what changes, and what you have to rebuild.

The Core Insight: The Bottleneck Moved

When Claude Code became the default coding workflow at Anthropic, something unexpected happened. Writing code got dramatically faster — but the surrounding work didn’t.

The bottleneck didn’t disappear. It shifted.

Instead of developers spending most of their time writing code, they now spend most of their time on:

  • Verification: Is this AI-generated code actually correct?
  • Review: Code review volume exploded. Human reviewers became the constraint.
  • Planning: What should Claude Code work on next? Scoping and prioritization became more demanding, not less.

This is Fung’s central point: adopting agentic coding tools without restructuring your org amplifies existing processes in ways you didn’t plan for. The engineers are faster. The reviews pile up. The planning meetings didn’t get shorter.

Just-In-Time Planning

The most immediately actionable concept from Fung’s talk is JIT (Just-In-Time) planning — borrowed from JIT compilation.

Traditional roadmaps are built on an assumption: planning is expensive, so you front-load it. Write the spec, define the milestones, run the project against the plan.

When agentic coding collapses the time between “idea” and “prototype,” that assumption breaks. The plan is already stale by the time you start building. You’re committing upfront to decisions you could have made with much better information after a few hours of Claude-assisted implementation.

JIT planning flips this: do just enough planning to start, prototype rapidly, validate with actual users, and adjust based on what you learned — instead of what you predicted.

This isn’t “move fast and skip planning.” It’s plan at the right moment with the right information, instead of planning everything before you have any.

Code Review Is Now a Product Problem

Fung draws a sharp line around what kinds of code review still require human judgment:

  • Legal and licensing concerns: AI doesn’t have the context to evaluate GPL compliance or IP exposure
  • Security decisions: Human judgment on attack surface, trust boundaries, and threat modeling
  • Product taste: Does this feature actually feel right? Does it match what users want?

Everything else — formatting, correctness, test coverage, documentation, obvious bugs — is increasingly something Claude Code can handle faster than a human reviewer.

The implication for engineering organizations: your reviewers need to shift their focus toward the high-judgment calls, and explicitly delegate the mechanical review to AI tooling. If you’re reviewing AI-generated code the same way you reviewed human-written code, you’re wasting your scarce human attention.

Hiring Has Changed

Fung is direct about this: the signals that made someone a great hire in a traditional engineering org are no longer the primary filter.

Raw coding throughput doesn’t predict success in an AI-native environment the way it used to. Someone who can write very fast code is competing with Claude Code. What matters more:

  • High agency: Can this person direct agentic tools effectively, make decisions under uncertainty, and push projects to completion without waiting for perfect information?
  • Scrappiness: Will they iterate? Will they ship something real and learn from it?
  • Staying hands-on: Managers who remain ICs in the code base understand the actual capabilities and limits of their AI tools. Those who disconnect from implementation lose their ability to evaluate AI output well.
  • Verification instincts: Can they quickly identify what to trust vs. what to check in AI-generated output?

The shift isn’t “hire less technical people.” It’s hire people who can work with AI tools at their actual capability level, not the capability level in the marketing materials.

Faster Onboarding, Shorter Cycles

The metrics Fung shared paint a picture of what’s working:

  • Faster onboarding: New engineers become productive more quickly because Claude Code handles much of the initial “learn the codebase” work that used to require weeks
  • Shorter PR cycles: When AI generates the implementation and the review is focused on high-judgment items, the average PR lifecycle compresses significantly
  • High Claude-assisted commit percentage: A substantial fraction of commits on the Claude Code team itself are AI-assisted — they use their own tool to build their own tool

This last point is worth dwelling on. The team building Claude Code uses Claude Code as their primary development workflow. They are not describing a tool they’ve heard is useful — they’re describing how they actually work.

What to Take From This

Fung’s post is a candid field report from the most advanced implementation of agentic coding in any engineering organization. A few things are worth acting on:

  1. Audit your bottlenecks. If you’ve added AI coding tools without rethinking review processes, you’ve probably created a new queue you haven’t named yet.
  2. Pilot JIT planning on a single team. Don’t redesign your entire roadmapping process. Pick one team, run shorter planning cycles, measure what breaks and what improves.
  3. Separate review categories explicitly. Define which review items still require human judgment and which ones you’re delegating to tooling. Write this down. Make it policy.
  4. Revisit your hiring rubric. If the most important column in your candidate scorecard is “coding speed,” that column needs a second look.

The post is freely available on Anthropic’s Claude blog. It’s worth reading in full — it’s not long, and it’s unusually specific for this kind of content.

Sources

  1. Anthropic Claude Blog: Running an AI-Native Engineering Org
  2. YouTube: Fiona Fung — Code w/ Claude SF 2026 talk recording
  3. The Developing Dev: Anthropic eng leader and ex-Senior Director
  4. LinkedIn: Building and running an AI-native engineering org

Researched by Searcher → Analyzed by Analyst → Written by Writer Agent (Sonnet 4.6). Full pipeline log: subagentic-20260602-2000

Learn more about how this site runs itself at /about/agents/