Descriptions:
Katia Gil Guzman and Vaibhav Srivastav from OpenAI’s Developer Experience team present a comprehensive, demo-driven look at Codex — OpenAI’s software engineering agent — at the AI Engineer conference. The session opens with a walkthrough of the model progression powering the system: GPT 5.2 and 5.3 Codex variants, the Cerebras-partnered GPT 5.3 Codex Spark for speed, the current flagship GPT 5.4, and the recently released GPT 5.4 mini and GPT 5.4 Nano designed for lightweight sub-agent workloads. Codex surfaces span a dedicated app, IDE extensions, CLI, Slack, and GitHub integrations.
Gil Guzman demonstrates the plugin ecosystem, including a game studio plugin that generates a complete playable platformer — sprites, mechanics, and all — from a single natural-language prompt with zero manual asset creation. She also shows web app plugins and the automations system for scheduling recurring agent tasks. A key demo involves Codex building a full game using its image generation skill, Playwright for interactive testing, and the game studio plugin in combination.
Srivastav then details the sub-agents architecture: Codex can now spawn specialized lightweight agents (GPT 5.4 mini or Nano) to handle parallelizable sub-tasks, enabling concurrent feature development, testing, and code review within a single workflow. He presents Codex’s code review capability as a quality gate for the reality that engineers can no longer manually review every line of agent-generated code — positioning automated review as an essential layer in the agent-assisted development loop.
📺 Source: AI Engineer · Published April 29, 2026
🏷️ Format: Hands On Build
