---
id: "lib-2026-05-28-openai-replay-2026-how-openai-uses-codex-to"
type: "library"
title: "OpenAI @ Replay 2026 | How OpenAI Uses Codex to Change How We Build"
lang: "EN"
org: "Temporal"
source_type: "Talk"
date: "2026-05-28"
week: "2026-W22"
url: "https://nosov.ai/library/lib-2026-05-28-openai-replay-2026-how-openai-uses-codex-to/"
original_url: "https://www.youtube.com/watch?v=NjaX4qt-O1Y"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "OpenAI Codex developer-experience lead ('Dom') describes internal practice at Temporal's Replay 2026; concrete and first-hand on context, verification and review, but a vendor talk with self-reported figures and product launches mixed in."
summary_basis: "transcript"
original_minutes: "47"
tags: ["enablement", "lecture"]
---

# OpenAI @ Replay 2026 | How OpenAI Uses Codex to Change How We Build

Tier B (Useful, limited): OpenAI Codex developer-experience lead ('Dom') describes internal practice at Temporal's Replay 2026; concrete and first-hand on context, verification and review, but a vendor talk with self-reported figures and product launches mixed in.

Summary based on the full transcript.

## In short

A 47-minute talk with Q&A at Temporal's Replay 2026 (May 2026) by an OpenAI engineer working on Codex developer experience. His thesis: code is no longer the bottleneck; quality depends on three things a team must set up for agents, as it would for a new hire: context, validation and verification. He walks through how he and OpenAI teams do each, with live examples.

## Key numbers

- 80%: OpenAI staff, including sales, marketing, finance and comms, said to use Codex weekly
- 100%: OpenAI pull requests run through Codex code review on GitHub before human review
- at least 50%: target speed-up for a docs build given to Codex as a measurable 'goal', reached in about two hours
- 23 minutes: end-to-end time for a Slack-to-PR task, of which six or seven minutes were CI

## What it reports

- Context: Codex is connected to Linear, Sentry, CI, Figma, Gmail, calendar, Slack and Notion, so "vague prompts" (find the thing a colleague asked me for) can be resolved. Codex can be tagged directly in Slack, Linear and GitHub. "Memories" and an opt-in screen-watching feature ("Chronicle") let it learn where information lives.
- Validation: agents use the same tools as engineers (compilers, formatters, tests, linters). A prose linter (Vale) enforces the docs style guide. Tool speed matters more in agent loops, so the team moved to faster native tools (oxfmt, TypeScript 7 preview). Parallel agents need isolated environments; UI work needs screenshots, accessibility tools or computer use.
- Error messages from linters and CI should tell the agent how to fix the problem.
- Review: every PR first goes through Codex review, tuned to flag only important issues; humans review once checks pass. Teams can raise the bar (typos count as P0/P1 in docs). Recurring misses are fed back into review rules or validation tools.
- CI and previews: "babysitting" skills let Codex watch its own PR until CI passes; deploy previews or recorded videos let reviewers check changes without running the branch, even from a phone. Subject-matter experts remain code-owner gates.
- Restraint: build, test internally, debate, ship, and delete features readily; the 3-month-old Codex app has had several large refactors without code freezes.
- Trends: proactive agents, more independent long-running work, and parallelisation in cloud environments. A "goals" feature runs until a measurable goal is met; one user's run lasted about three days.
- Q&A: specs and test suites become the source of truth; conflicting instructions are a bigger problem than too much context; keep AGENTS.md short.

## How they got it

Personal and team practice at OpenAI, shown through screenshots and live demos. No measured outcomes beyond anecdotes.

## Limits

- Speaker works on the product and has "unlimited tokens"; cost constraints are waved away.
- Several features are previews or launches announced in the talk.
- Usage figures are internal and unverified.
- Auto-captioned; product names are partly mis-transcribed.

## Why it matters for companies

The context-validation-verification checklist applies to any team deploying agents, in or outside software: can the agent reach the information it needs, can it check its own work, and who confirms the result. Two practices transfer directly: run automated review before human review, and turn every repeated miss into a rule or check. Connecting agents to email, chat and screen activity raises access, privacy and works-council questions that need answers before copying OpenAI's setup.

Original: https://www.youtube.com/watch?v=NjaX4qt-O1Y