“Engineers are becoming sorcerers” | The future of software development with OpenAI's Sherwin Wu

95% adoption with full AI code generation

At OpenAI, 95% of engineers use Codex daily and 100% of PRs are reviewed by AI, with close to 100% of code being AI-generated first before human review.

70% increase in PR output

Engineers using Codex open 70% more pull requests than non-users, with the productivity gap widening as proficient users learn to leverage the tools more effectively.

This is the floor, not the ceiling

As VP of Science Kevin Weil notes, current models represent the worst they will ever be, with capabilities improving exponentially from this baseline.

From coder to tech lead

Software engineers are evolving into managers overseeing fleets of agents rather than writing code, juggling 10-20 parallel AI threads simultaneously.

The sorcerer's apprentice dynamic

Like Mickey Mouse in Fantasia, engineers cast high-leverage 'incantations' but must remain vigilant to prevent agents from going off-rails without proper supervision.

SICP prophecy fulfilled

The 1980 MIT textbook 'Structure and Interpretation of Computer Programs' described programming as sorcery, a metaphor now literal as engineers cast spells that generate complex software automatically.

The no-escape-hatch experiment

An internal OpenAI team maintains a codebase written entirely by Codex without the option to manually code, forcing them to solve agent failures through prompting and context rather than human intervention.

Underspecification is the primary enemy

When AI agents fail, the problem is almost always insufficient context or underspecification, requiring engineers to encode tribal knowledge into documentation, code comments, and repository structure.

Context engineering over code writing

Success requires creating comprehensive markdown files, skill definitions, and architectural documentation that allows agents to understand intent without human translation.

Build for tomorrow's models, not today's

Teams should architect for where models are heading rather than current capabilities, as rapid evolution means today's scaffolding gets 'eaten for breakfast' by next-generation models.

The one-person billion-dollar startup era

Second and third-order effects of solo unicorn companies will include ecosystems of hundreds of small startups building bespoke B2B SaaS tools, potentially triggering a golden age of business software.

Customer listening has diminishing returns

In AI's current velocity, listening to existing customer needs can mislead builders because models disrupt themselves so quickly that today's solutions become obsolete before deployment.