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

The Transformation of Software Engineering at OpenAI

At OpenAI, the adoption of AI in the engineering workflow is nearly universal. 95% of engineers use Codex on a daily basis, and 100% of pull requests are reviewed by Codex before being merged. This has led to a tangible shift in productivity and the nature of the engineering role itself.

• Productivity Gap: Engineers who heavily use Codex are opening 70% more PRs than those who don’t, and this gap is widening. This suggests that proficiency with AI tools creates a significant and growing advantage.
• Accelerated Code Reviews: With Codex providing a “smart second pair of eyes,” code review times have been drastically reduced from 10-15 minutes to just 2-3 minutes. For many small changes, engineers trust Codex’s review without needing a human counterpart.
• Engineers as Sorcerers: The role of an engineer is evolving from a writer of code to a manager or “tech lead” of AI agents. Many engineers manage fleets of 10 to 20 parallel Codex threads simultaneously. Sherwin Wu draws a parallel to the classic programming textbook Structure and Interpretation of Computer Programs (SICP), which described programmers as “sorcerers” casting “incantations” (code). With modern AI, this metaphor has become literal. However, it comes with the challenges depicted in “The Sorcerer’s Apprentice,” where unskilled use of powerful tools can lead to chaos, highlighting the need for skill and oversight.

A 100% AI-Written Codebase: Learnings from the Frontier

To push the boundaries of AI-driven development, an internal team at OpenAI is conducting an experiment to maintain a codebase that is 100% written by Codex. This team intentionally has no “escape hatch” to write code manually when the agent gets stuck. This constraint has yielded a crucial insight:

When an AI coding agent fails to perform a task, the root cause is almost always a lack of context.

The solution is not to manually intervene but to enrich the agent’s environment. The team focuses on encoding tribal knowledge directly into the codebase through:

• Detailed code comments.
• Comprehensive documentation in Markdown files (skills.md).
• Clearer code structure.

By improving the available information, the agent can overcome obstacles on its own, demonstrating a new paradigm for unblocking AI-driven development.

The Changing Role of the Engineering Manager

While the manager’s role has changed less than the engineer’s, several key trends are emerging:

1. Amplify Top Performers: AI tools disproportionately benefit high-agency, top-performing engineers, making them exceptionally productive. Wu advises managers to spend the majority of their time—more than 50%—with these top performers, focusing on unblocking them and ensuring they have the resources to push the boundaries.
2. Increased Leverage and Team Size: AI tools that provide organizational context (e.g., a chatbot connected to internal Notion, Slack, and GitHub data) give managers higher leverage. This allows them to effectively manage much larger teams than the traditional 6-8 direct reports, as they can quickly gather context, track progress, and even get assistance with tasks like writing performance reviews.
3. The Manager as a Support System: Drawing from The Mythical Man-Month, Wu describes his management philosophy as creating a support system for his engineers, akin to a surgical team supporting a surgeon. The manager’s job is to look around corners, anticipate needs, and proactively remove organizational blockers so the engineer can remain focused and productive.

Building in a Rapidly Evolving AI Landscape

The pace of change in AI means that traditional product development strategies, like blindly following customer feedback, can be misleading.

“The models will eat your scaffolding for breakfast.”

This phrase captures the idea that the complex tooling and frameworks (the “scaffolding”) built to overcome the limitations of today’s models will likely become obsolete as future models become more capable. This is a direct application of “The Bitter Lesson” in AI research to product development. For example, the focus on complex vector search strategies and rigid agent frameworks in 2023 is already giving way to simpler, model-driven approaches that use tools and file-based context.

The key advice for builders is:

“Make sure you’re building for where the models are going and not where they are today.”

Build a product for a capability that is perhaps 80% of the way there. As the underlying models improve, the product will “click” and become truly magical, leapfrogging competitors who built for the limitations of the past.

The Future of AI: What to Expect in 12-18 Months

Wu highlights two key areas of advancement to watch for:

1. Long-Duration Coherent Tasks: Models will evolve from handling tasks that take minutes to coherently executing tasks that span multiple hours or even a full day. This will unlock entirely new categories of products and applications.
2. Advanced Multimodality (Audio): Audio, especially native speech-to-speech models, is an underrated domain poised for significant breakthroughs. Given that much of the world’s business is conducted via audio, improvements here will have a massive impact on enterprise and service industries.

OpenAI’s Platform Philosophy

Addressing the common startup fear of being “squashed” by the platform they build on, Wu emphasizes that OpenAI is fundamentally an ecosystem and platform company. This approach is rooted in its core mission to ensure the benefits of AGI are distributed broadly to all of humanity, which the company knows it cannot achieve alone.

Startups should worry less about competition from OpenAI and more about finding product-market fit. The opportunity space is so vast that building a product people truly love is the most direct path to success.

• Tags:
• AI Agents
• Software Engineering
• Large Language Models
• Developer Productivity
• Feature
• OpenAI
• Codex
• Engineering Management