Slow down to speed up: so much has changed in 6 months’ time
- Source
- newsletter.pragmaticengineer.com
- Author
- Gergely Orosz
- Date
Context
One response to AI-generated code is to decide explicitly which changes still require human review. In his account of changing development practices, Gergely Orosz describes a tiered process on OpenAI's Codex team. AI reviews all changes, some can proceed with that review alone, and more important changes also require a person. The reported mechanism allocates human according to the significance of the change. It shows how adopting coding agents can alter the approval process as well as the way code is produced, making the boundary for human intervention a deliberate engineering decision.
- Claude Code's creator runs roughly five agents in parallel and ships 20-30 pull requests a day. Inside Anthropic, written product requirement docs have been replaced by prototypes, and Claude Cowork was built in ten days.
- Average pull request size is up 3x in 18 months. Review load therefore grows faster than PR count, and the bottleneck moves from writing code to having anyone left who can read it.
- Meta's cuts landed unevenly: Instagram's design team lost 44% of headcount, developer documentation and support lost 95%, and senior integrity staff were reassigned to AI data-labeling work.
- DHH reversed his own position that AI should not write code directly. His resistance was about model quality — he used to spend longer rewriting the output than writing it himself, and that arithmetic flipped.
- attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
“The models released in November 2025 elevated agents to being genuinely useful. We"”
“Meta has been destroying its engineering org, and an obsessive focus on AI seems to be one reason for it.”
Checking sign-in…
Loading comments…