OpenAI Codex Masterclass — Vaibhav Srivastav & Katia Gil Guzman

April 29, 20261:01:5941,029 views

Summary

Vaibhav Srivastav (VB) and Katia Gil Guzman from OpenAI's Developer Experience team delivered a masterclass on OpenAI Codex at AI Engineer Europe 2026, covering the full software engineering agent ecosystem — from underlying models to plugins, automations, subagents, and bleeding-edge experimental features.

The session opened with a model overview: Codex runs on a progression of frontier models — GPT-5.2, GPT-5.2 Codex, GPT-5.3 Codex, GPT-5.3 Codex Spark (deployed on Cerebras for extreme speed), and the latest GPT-5.4 (with GPT-5.4-mini and GPT-5.4-nano variants for short-running and subagent tasks). A unified agent harness wraps all models to manage tool execution, environment setup, and safety. Performance improvements include WebSocket-based token delivery providing approximately 1.75x faster tokens, and a Fast Mode that doubles this to an effective 2x speed gain. As of the day of the talk, Codex surpassed 3 million weekly active users — more than triple the number from January 2026.

Katia demoed the Codex app's plugin system, which bundles skills (reusable instruction packages), apps (third-party service connections), and MCP servers into single-install workflows. Featured plugins included a Game Studio plugin (combining Imagen for generating sprites and Playwright Interactive for headless browser debugging), a Google Drive plugin (demonstrated by scanning 57 YAML-based Codex meetup entries in a repo and writing them to a Google Sheet in under two minutes), and web/app development bundles. Automations — cron-style background jobs within the Codex app — were shown for Slack message triage (daily 9 a.m. summaries bucketed by topic) and Gmail filtering for time-sensitive replies.

VB presented Codex's code review capabilities, noting that 100% of pull requests across all OpenAI repositories — including those by Greg Brockman — are reviewed by Codex code review as a first pass by default. The review system spins up a dedicated Codex thread with a specialized review system prompt that contextualizes diffs against the full codebase, identifying second-order effects beyond the immediate changeset. Reviews are prioritized as P0/P1/P2 findings. A Cloud Code plugin (co-shipped with colleague Dom) also brings the same review capability into Claude Code sessions.

Subagents were covered in depth. The system allows a master task to be decomposed into parallel, independent subtasks handed off to named agent instances. By default, three personas ship: a general-purpose fallback agent, a worker (execution-focused), and an explorer (read-only traversal). VB demonstrated spinning up 20 subagents to review 45 persona YAML files in a Codex Agents repo, with Codex automatically entering plan mode and partitioning files across agents. Custom subagents can specify model (e.g., GPT-5.3 Codex Spark for read-only exploration), reasoning effort, sandbox mode (read-only vs. write), and granted MCP server access (e.g., Sentry, Linear). A PR Explorer persona example used GPT-5.3 Codex Spark in read-only mode to trace execution paths without proposing fixes.

Bleeding-edge experimental features included: Guardian Approvals (accessible via /experimental), a safety system that spins up a subagent to evaluate whether each privileged action needs human intervention, reducing approval fatigue without full YOLO mode; Hooks (three supported: session start, per-tool-use, and stop), demonstrated with a keep_going.py stop hook that prompts Codex to continue long-running tasks autonomously; Personality settings for tone customization; Codex Security for commit-by-commit vulnerability scanning and patching; and native Windows sandbox support, noted as the first in the industry among competing coding agent harnesses.

Key Takeaways

•Codex is a full software engineering agent — not just a code writer — capable of running commands, tests, and exploring codebases via a unified agent harness.
•Spin up subagents with explicit model assignments: use GPT-5.4-mini or GPT-5.4-nano for short-running subagent tasks to reduce cost, and reserve GPT-5.4 for complex long-running work.
•Always set review and security subagents to read-only sandbox mode to prevent unintended execution during analysis.
•Guardian Approvals (enable via /experimental) replaces full YOLO mode with a subagent-based privilege evaluator, significantly reducing approval fatigue on long-running tasks.
•A stop hook with a keep_going.py script enables fully autonomous long-running Codex sessions without manual re-prompting.
•100% of OpenAI's internal pull requests are first-pass reviewed by Codex code review — adopt this as a baseline before any human review cycle.
•Package repetitive workflows into plugins (skills + apps + MCP servers) rather than configuring them manually each session.
•Ask Codex to scan your .sessions directory and auto-generate recommended automations and subagent personas based on your actual usage patterns.
•For visual app and game development, the Playwright Interactive skill + Imagen skill combination enables Codex to generate assets and debug the running UI in a headless browser autonomously.
•Give individual subagents scoped MCP access (e.g., Sentry for error reports, Linear for issue triage) rather than broad permissions, to keep tasks isolated and safe.
•Use GPT-5.3 Codex Spark (Cerebras-hosted) for blazingly fast read-only exploration subagents where inference speed matters more than frontier capability.

Transcript

Hi everyone. Thank you for being here. So, today we're going to talk about Codex. My name is Katya. Katya Gribuzina, and I'm with VB. Uh we are both uh working in the developer experience team at VB. based in London, and so our role is really to help developers build and get the most out of our products, uh including Codex. And so, today we're going to start with a quick Codex overview, uh just so we know how many of you here are using Codex. Can you raise your hands? Yay! Okay, cool. So, we we'...