Inside OpenAI’s agentic software factory

gist

From the free public preview of Gergely Orosz's OpenAI visit (sections 1–3 and the opening of 4; the rest is paywalled), Codex and ChatGPT Work have become the company's shared harness: non-engineers adopted without mandate, IDE usage fell as long-running threads rose, and CI/CD load jumped about 10x in months. OpenAI runs an agentic software factory—context gathering, implement, CI babysitting, specialist reviews, risk-tiered merge, agentic deploy with per-change dashboards, Perf Factory feedback, and Sevbot incident assist—built around that harness.

ideas

• Codex took over without a top-down mandate. Non-engineering orgs moved from near-zero to ~90% usage as the Mac/Windows app and ChatGPT Work shipped; long-running /goal threads and role-specific plugins beat a blank generic agent.
• The IDE and classic PRs are under pressure. Desktop Codex won the between-CLI-and-IDE bet as IDE usage fell after January; PR volume is hockey-sticking and reviews are being reimagined as multi-lens agent reviews plus risk-tiered human gates.
• A software factory is closed agentic loops. Humans set outcomes; Codex pulls repo/Slack/Notion/Datadog context, implements, greens CI (including a perf harness), specialist agents review, deploy agents handhold rollouts and build dashboards, Perf Factory feeds regressions back, and Sevbot assists incidents without autonomous mitigation yet.
• Harness wiring and awareness dominate adoption. Internal Codex is deeply plugged in; capability overhang gave way to awareness overhang as teammates discover Slack/Airtable/onboarding uses by word of mouth.
• Preview boundary. This digest covers only the free public preview. Sections on tooling practices, billion-user infra, API reliability, and how the SE job is changing were not captured beyond the paid gate.

quotes