Codex takes over OpenAI: inside an agentic software factory

A visit to OpenAI headquarters reveals a sharp shift in how work gets done. In four months, nearly every non-technical employee — finance, recruiting, legal — went from zero Codex usage to 90% weekly adoption. The tool launched as a Mac app in February, Windows in March, and ChatGPT Work in July. It became the default without a top-down mandate. "The big theme of the last few months is that everything is a code agent now," said Andrew Ambrosino, head of the desktop team. "Whether visible code is your output or not, agents write your deliverables."

Token-usage graphs show a steep jump in January alongside a collapse in traditional IDEs. Pull requests and code reviews are being rethought from the ground up: when the agent writes the code, the human review process changes fundamentally. Joe Gershenson, who leads the core agents team, said legacy practices no longer match the new workflow.

OpenAI has built a "software factory" around automated feedback loops. A flagship example is Perf Factory, a system that monitors production and spins up Codex agents to fix performance issues automatically. Manually built internal tools are being replaced by Codex, which engineers now prefer for debugging over purpose-built alternatives. The efficiency of the harness — the infrastructure that runs the agents — has become critical.

Engineering for a billion users: on the infrastructure side, the rule is buy first, bring in-house later. Geographic distribution, capacity planning, and load challenges are handled pragmatically. On the API side, CPUs became the bottleneck, so the team moved to intentional slow rollouts to improve reliability. Engineers such as Ahmed Ibrahim and Steve Coffey describe a shift from narrow specializations to judgment and agency as the core skills. The result: migrations and rewrites once deemed "impossible" now succeed with one or two engineers.