AI
AI Adoption Reshapes Engineering, Not Just Code
At Google, 75% of all code is written by AI. DORA research across multiple companies found that AI adoption increases individual productivity while sometimes decreasing team-level outcomes. The engineers who use AI the most spend more time coding, ideating, and collaborating. They are more active, not less. The software engineering role is not vanishing. It is evolving toward higher-altitude work where intent and system design replace line-by-line coding.
Five Patterns Emerge in AI-Native Engineers
Nicole Forsgren identified five patterns in Google’s highest-performing AI-native engineers. First, they operate at higher altitudes, combining engineering knowledge with business and user context. Second, they shift left on intent, writing specification files that capture decisions before agents execute. Third, they design environments and guardrails rather than live coding. Fourth, they work in small cross-functional pods with less communication overhead. Fifth, they maintain a scientific mindset, experimenting and codifying learnings back into the system.
Core Engineering Skills Control the AI
Andrew Macvean shared a story from Google’s search organization where PMs ship features to production without writing traditional code. The platform that enables this requires deep engineering expertise in system architecture, scalability, and reliability. AI can generate code fast, but verification becomes the bottleneck. Engineers must set feedback loops to catch issues reliably. The saying inside Google: delegate tasks, not judgment. Practical tips include re-implementation as a learning tool, walkthroughs of alien code, and maintaining skill files for agents.
From Prompting to Orchestration
The shift from AI enhancement to AI-native engineering means moving from conductor to orchestrator. Google’s top engineers manage teams of asynchronous agents with specific context windows and responsibilities. Code review agents evaluate functionality, reliability, security, and maintainability. Shepherding agents guide changes through CI/CD pipelines. Risk assessor agents flag high-risk changes for human review. Agent journaling helps engineers improve instructions and the environment around agents. TensorFlow migrations at Google use a three-agent architecture: planner, orchestrator, and coder.
Adjacent Skills Prevent Quiet Catastrophic Debt
AI acts as an amplifier of existing strengths and a mirror for weaknesses. Google has not seen any measurable increase in outages despite more AI-written code. Tiered risk environments contain the blast radius while enabling faster feedback loops. Adjacent engineering skills like deployment infrastructure, security, and compliance become critical. Engineers act as value translators, turning user needs into precise requirements. Spec-driven development treats the spec as the ultimate deliverable, with goals and constraints documented before any agent executes.
Leaders Must Build the System That Enables T-Shaped Growth
You cannot mandate a T-shaped developer inside a broken system, Macvean said. Leaders should redefine productivity metrics away from pull request throughput toward business outcomes. They must protect productive struggle by carving out time for engineers to learn tools and build mental models. They need to foster radical psychological safety so teams can experiment and fail without blame. 10x output cannot come with 10x cognitive load. A bad system will beat a good person every time.
Notable Quotes
code is written by AI. Andrew Macvean · ▶ Watch (1:53)
AI is an amplifier and it is a mirror. Nicole Forsgren · ▶ Watch (30:19)
acting as value translators Nicole Forsgren · ▶ Watch (35:41)
foster radical psychological safety. Nicole Forsgren · ▶ Watch (42:34)
Key Takeaways
- Google’s top AI-native engineers are more active, not less, spending more time on ideation and collaboration.
- Specification files and agent skill files replace code as the first-class outcome in AI-native engineering.
- AI amplifies existing strengths and exposes weaknesses; fragmented tooling generates technical debt faster.