Harness Engineering: Leveraging Codex in an Agent-First Environment

For the past 5 months, OpenAI ran an experiment building internal beta software using Codex under the constraint of 0 lines of manually written code. 3 engineers operated Codex to process about 1,500 PRs, completing a 1 million-line-scale codebase in 5 months.

A Fundamental Shift in the Engineer's Role The engineers' primary work shifted from directly writing code to designing environments, specifying intent, and building feedback loops so that agents could work reliably. They focused on breaking large goals down into small building blocks the agents could act on, and on building processes that let agents review and fix their own changes.

Optimizing the Development Environment for Agents

• Application Readability: Designed so agents can directly read and verify UI, logs, and metrics. They use DOM snapshots and screenshots via the Chrome DevTools Protocol, and query logs and metrics with LogQL and PromQL to reproduce bugs.
• Context Management: Instead of massive instruction manuals, they use a short AGENTS.md as a map, and manage knowledge through a structured docs/ directory. This lets agents navigate to the information they need without being overwhelmed by excessive information.
• Execution Plans: Complex tasks are stored as version-controlled execution plans, allowing agents to grasp task progress and decision logs without external context.

Remaining Challenges To achieve high throughput, they minimized merge gates and adopted a post-hoc correction approach, but maintaining long-term architectural consistency and how to effectively encode human judgment into the system remain ongoing learning challenges.