A New Paradigm: Humans Steer, Agents Execute
Within five months, an OpenAI team completed a challenge: developing and launching an internal product with code written entirely by an AI Agent. The project started in August 2025 from an empty repository, driven by an engineering team of initially just 3 people. By early 2026, the repository had accumulated about 1 million lines of code, covering everything from application logic to CI/CD configurations, and had merged approximately 1,500 Pull Requests generated by the AI. The product is now used daily by hundreds of internal users and has entered an external alpha testing phase. The entire process strictly followed the principle of “humans write no production code.” The estimated development efficiency is 10 times that of traditional manual coding.
Redefining the Engineer’s Role: From Coding to Design
In this experiment, the core responsibility of engineers was no longer writing code but shifted to higher-level tasks like system design, environment construction, and feedback loop optimization. The team’s primary task was to build a “standardized” and “readable” work environment for the AI Agent (Codex). For example, by integrating with the Chrome DevTools Protocol, they enabled the agent to directly manipulate the UI and reproduce bugs. By exposing LogQL and PromQL query interfaces, they allowed the agent to understand application logs and performance metrics. This “scaffolding” work to empower the AI was key to achieving effective human-agent collaboration.
Enforcing Architecture with a Structured Knowledge Base
To enable the AI Agent to handle complex tasks and maintain system architecture consistency, the team abandoned single, monolithic instruction files. Instead, they built a structured docs/ directory as the “single source of truth.” A concise AGENTS.md file serves as a high-level map, guiding the agent to consult more detailed design documents, architectural standards, and execution plans when needed. More importantly, the team used custom linters and CI jobs, also generated by Codex, to mechanically enforce architectural rules, such as strict dependency directions and module layering. These coded “invariants” effectively prevented the architectural decay that can occur during rapid iteration.
Towards Autonomy and Continuous Maintenance
As the system matured, the AI Agent’s level of autonomy significantly increased. It can now complete end-to-end bug fixes from a single prompt, including reproducing the issue, writing code, validating the fix, opening a PR, and responding to review comments. To combat the potential “entropy” or technical debt introduced by AI-generated code, the team also deployed periodically running “documentation gardener” and “garbage collection” agents. These automated tasks scan and refactor code that deviates from best practices, ensuring the ongoing health of the codebase. This practice demonstrates that the future challenges of software engineering are shifting from writing code to designing environments and systems that can effectively guide and control AI agents.
