Inside OpenAI's Agentic Software Factory: Codex as Infrastructure

Non-Engineers Adopt Codex First

OpenAI internal data shows non-engineering departments (finance, recruiting, legal) went from near 0% to 90% Codex adoption in four months. The Mac app launched in February, Windows in March, and ChatGPT Work (built on Codex harness) in July. Early versions showed code directly, yet non-technical staff used it for research, presentations, and spreadsheets — caring about results, not code.

The inflection point came in April–May when OpenAI added the /goal command: set a target and the agent runs until completion. Adoption jumped from 60% to 90%. Desktop lead Andrew Ambrosino observed people spending days on a single thread, letting the model run long tasks that often spawn sub-agents, reducing the surface area humans must manage.

Productivity engineering lead Akshay Nathan framed this as a shift from "capability overhang" (model ability exceeded product) to "awareness overhang" (product is ready but users don't know its full potential). Teams began packaging workflows as plugins; Ambrosino noted: "If you build a product that does everything, teams need a way to make it their own. You can't give everyone an empty box."

New engineers are onboarded by asking Codex, which has access to virtually all internal systems — Git, Slack, Notion, Databricks, Datadog. All documentation lives in the codebase, giving agents more context than any human. Some internal skills are maintained by Codex itself. The internal version far outpaces the external one due to deep system integration, mirroring Ramp's Insight AI approach: deep integration beats generic capability. Dependency is so high that a small outage triggers Slack alerts faster than automated monitoring.

IDE and Pull Requests Are Fading

Codex team initially debated a desktop app, fearing it would sit between CLI and full IDEs like an unused iPad. When Antigravity (a VS Code fork) appeared in November, they bet that stronger AI agents would make IDEs less important. They were right: IDE usage dropped steadily after January 2025. Codex desktop later added in-app file editing (June), but the direction was set.

PR volume exploded hockey-stick style; CI/CD load grew ~10x in six months — a scale that typically takes 2–3 years. VP of Applied Infrastructure Venkat Venkataramani said: "Every month we hit a new infrastructure bottleneck. Just as capacity seems enough, the model unlocks a new capability and the bottleneck moves." Their response: rethink PR and code review entirely. "The way we do code review today makes less and less sense, and so do PRs."

Now multiple agents review in parallel, each acting as a domain expert (cloud infrastructure, security, etc.) with full codebase and documentation access. Changes are risk-tiered; low-risk PRs can be auto-approved by agents, only high-risk changes require humans. Gergely initially doubted whether a prompted "cloud expert" agent would differ from a generalist, but realized each agent's focused context window enables deep, specialized review.

The Agentic Software Factory Pipeline

Orosz likens the system to a "dark factory" — fully automated, no humans needed. The pipeline:

Define Goal. Engineers or PMs describe problem and desired outcome. Venkat notes engineers are becoming more like PMs: judgment, prioritization, and taste matter more than writing code.

Gather Context. Codex taps Git, Slack, Notion, Databricks, Datadog, internal logs. Documentation in code gives agents context far beyond human capacity.

Implement Code. Codex makes iterative changes until the goal is met, verifying software works.

Build, Test, CI. Agents build, run tests, fix failures, open PRs, and guard them until green. A perf harness routes potentially risky PRs to A/B evaluation.

Agent Code Review. Multiple domain-expert agents review in parallel, risk-tiered. Human review cannot keep pace with output velocity.

Agent Deployment. Each change gets a dedicated agent tasked with "shepherd this change safely to production." The agent reads code, finds feature flags, defines success/failure signals, builds per-deployment monitoring dashboards. Long-term goal: a "per-change autonomous SRE."

Perf Factory. Automated performance optimization: agents scan alerts, deduplicate, identify true latency regressions, root-cause, and propose fixes — extending the workflow from deployment to continuous optimization.

Sevbot. Incident-response agent that gathers context, suggests mitigations, answers questions, but does not execute actions. Dream: handle routine nightly incidents without waking humans. On-call not yet eliminated.

The Glaring Bottleneck: App Store Review

Everything accelerates except Apple/Google app store review. ChatGPT engineering lead Sulman Choudhry (ex-Facebook) recalls Facebook using feature flags to compress mobile releases from monthly to biweekly to weekly. Codex writes code in minutes; getting it to users' phones takes days. "If software can be written in minutes but takes days to reach phones, that contradiction becomes increasingly absurd." App Store review has barely changed in 18 years since 2008; Apple still blocks meaningful experience updates without review.

Engineering Roles Are Shifting

All seven interviewees highlighted the same trend: specialization blurs, judgment and agency outweigh coding. Previously "impossible" rewrites and migrations now take one or two engineers plus an agent. OpenAI embedded domain experts (slide design, financial reporting) into the ChatGPT Work engineering team because model "taste" in those domains exceeds developers' — developers can't judge output quality, so real experts define what good looks like. Gergely notes this is a decades-old best practice rediscovered every few years.

Key Data Points (OpenAI Economic Research, June 2026)

By May 2026, 80.6% of sampled users made at least one Codex request estimated to replace >30 minutes of human labor; 70.2% >1 hour; 25.6% >8 hours.

Codex accounts for 99.8% of OpenAI's internal weekly output tokens.

Heavy users (99th percentile) run >60 hours of agent turns per day via massive parallelism.

Since August 2025, non-developer individual users grew 137x; organizational non-developer users grew 189x.

Over 25% of business-department (finance, marketing, ops) work done via Codex is coding/engineering tasks — agents are breaking role boundaries.

The verdict is clear: inside OpenAI, Codex is not a tool but infrastructure, and the workflow running on it is no longer traditional software engineering.