Inside OpenAI’s agentic software factory — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Codex adoption at OpenAI jumped from ~0% to 90% across non-engineering teams (finance, recruitment, legal) in roughly four months, with no mandate from leadership — driven largely by word-of-mouth and the Codex app's ability to handle long-running tasks
- OpenAI's "software factory" runs automated agentic feedback loops including Perf Factory, which monitors production and kicks off Codex agents to automatically fix performance issues, replacing what used to require dedicated human engineers
- PRs per engineer at OpenAI are growing "like a hockey stick" with roughly a 10x load increase on some systems in about six months — a level of growth that takes most companies two to three years, exposing bottlenecks across version control, CI/CD, and release processes
- IDE usage has declined since January when Codex usage started surging, and the Codex team considered but rejected forking VS Code, betting instead that "as AI agents get better, IDEs will matter less"
- ChatGPT Work (powered by the Codex harness) launched in July, following Codex app releases for Mac in February and Windows in March; a long-running goal-setting feature drove usage from 60% to 90% between April and May
- Domain experts are now embedded in ChatGPT Work engineering teams because models have become "smarter" than developers in domains like slide decks, spreadsheets, and business reports — devs can no longer channel "taste" into the harness in those areas
- OpenAI's internal Codex is significantly more advanced than its external counterpart because it's plugged into nearly every internal system, and the company's dependence on it is so total that minor outages trigger internal alerts before automated ones
Why it matters: OpenAI is now so dependent on a single Codex-powered harness that even minor outages halt work across the entire company, yet its CI/CD systems are absorbing roughly 10x load in six months — a fragility-versus-scale tension that will test whether the "agentic software factory" model holds under production stress, with implications for any company trying to replicate the pattern.
Ask SkimNews
