From syntax to orchestration.

Shifts

1. AI writes the code nowAI writes the code nowAI writes the code now75%

   Shift

   75%

   of new code at Google is AI-generated

   Up from 25% in late 2024. In eighteen months the job flipped from typing code to directing the models that write it.

   Source: Sundar Pichai, April 2026

2. The harness beats the modelThe harness beats the modelThe harness beats the model+13.7 pts

   Shift

   +13.7pts

   on Terminal-Bench - same model, new harness

   LangChain changed only the prompts, tools, and middleware around GPT-5.2-Codex and climbed 25 leaderboard places. The leverage moved to the setup, and the setup is engineering.

   Source: LangChain, harness engineering

3. Working is cheap. Secure is not.Working is cheap. Secure is not.Working is cheap. Secure is not.45%

   Shift

   45%

   of AI-generated code failed security tests

   Veracode tested 100+ models: the code ran, and nearly half of it shipped OWASP Top 10 flaws. Anyone can get an app running by evening - shipping it safely is still the job.

   Source: Veracode GenAI Code Security Report, 2025

Principles

1. Never ship what you can't explainNever ship what you can't explainNever ship what you can't explain19%

   Principle

   19%

   slower with AI - while feeling 20% faster

   In METR's randomized trial, experienced maintainers on their own repos lost time checking AI output and never noticed. If you can't explain the change, you can't tell whether it helped.

   Source: METR, July 2025

2. Taste doesn't automateTaste doesn't automateTaste doesn't automate66%

   Principle

   66%

   of developers fight AI answers that are almost right

   "Almost right, but not quite" is developers' top AI frustration. Spotting the quietly wrong answer - and deciding what's worth building at all - stays with you.

   Source: Stack Overflow Developer Survey, 2025