In his groundbreaking article titled “Harness Engineering,” Ryan Lopopolo from OpenAI discusses the innovative practices of OpenAI’s Frontier team, now the largest user of Codex. The team operates an extensive, 1 million-line codebase with zero manual code or human reviews. Lopopolo stresses that modern software development must utilize over 1 billion tokens daily—costing $2,000 to $3,000 to maximize efficiency. His team’s ambitious experiment aimed to build a test product without any human-written code, resulting in the creation of Symphony, a “ghost library” for a large-scale Codex agent system.
By focusing on optimizing for agents rather than human workflows, the Frontier team identified human attention as the bottleneck in AI-driven development. They keep build times under one minute, improving productivity through constant refinement. Ultimately, Lopopolo’s insights position coding agents as essential “teammates,” encouraging developers to embrace automation in their engineering practices for enhanced efficiency and innovation.
Source link