Recently, OpenAI engineer Ryan Lopopolo revealed the intensive usage of Codex by internal teams through a long article “Harness engineering”. This experiment not only produced a "ghost library" called Symphony, but also defined a new way of engineering work.
Core Experiment: Putting Yourself in a "No Human Coding" Shackles
Ryan set an exaggerated restriction at the beginning of the experiment: never write any code by hand.
From Pain to Breakthrough: In the early stage of the experiment, due to the model's difficulty in handling complex logic, the development speed was only one-tenth of that of humans.
Forcing Tool Evolution: To enable the Agent to complete tasks independently, the team had to develop an extremely refined
Exponential Efficiency: With the iteration of models like
Key Tactics: Keeping Build Time Within "One Minute"
Within
Inner Loop: The team continuously rebuilt the build system (from Makefile to Bazel and then to Nx), ensuring that the build time remained strictly within 1 minute.
Logical Decomposition: If the build slowed down, the Agent would automatically break down the build graph into finer parts until the complexity fell below the threshold, maintaining the Agent's "focus".
Paradigm Shift: People Became the Bottleneck of PR Review
When code could be easily parallelized, the real scarcity was "human attention".
Post-Merge Review: The team no longer conducted lengthy manual reviews before merging, but automated them. Most manual reviews occurred after the code was merged, used for quality checks and experience accumulation.
Observability Empowerment: Engineers' work is no longer fixing bugs, but providing
Experience "Distillation": The "tacit knowledge" in senior engineers' minds is written into Skill documents and tests, solidifying it as part of the system context.
Future of the Industry: "Ghost Libraries" and Internalization of Dependencies
Ryan proposed a striking idea: As token costs approach zero, software dependencies may gradually disappear.
Ghost Libraries: Developers just need to define a high-fidelity Spec, allowing the Agent to reassemble and implement it locally.
Internalization of Dependencies: For third-party libraries of medium to low complexity, the Agent can rewrite them inline into the repository, stripping out unnecessary code and building a fully self-controlled tech stack.
Conclusion: From "Co-pilot" to "Independent Teammate"
In the view of
