Harness Engineering: Leveraging Codex in the Agent-First Era

Over the past 5 months, an experiment was conducted to build a software product with not a single line of manually written code. Codex wrote all the code, including application logic, tests, CI configuration, and documentation, completing about 1 million lines of code using roughly 1/10 the time compared to the traditional approach.

In this process, the engineer's role shifted from writing code directly to designing the environment, specifying intent, and building feedback loops so the agent can perform tasks reliably. The engineer directs tasks through prompts, and development proceeds through a loop in which the agent reviews and revises its own code.

As development speed increased, the bottleneck shifted to human QA (quality assurance) capacity. To address this, the focus was on making the application UI, logs, and metrics directly readable and understandable by Codex, in order to increase the agent's autonomy.