Codex ran OpenAI DevDay 2025 (developers.openai.com)

At OpenAI DevDay 2025, Codex was the backbone behind demos, apps, booths and even arcade cabinets — used across stage presentations, the Apps SDK launch, and internal tooling to accelerate development, prototyping and iteration. Concrete wins included Codex implementing the VISCA camera protocol and building an MCP server to control venue lighting for a realtime demo; spinning up multiple single-file Phaser games in parallel for ArcadeGPT; converting a Streamlit fine-tuning demo into a Next.js + FastAPI app overnight; and automating docs by splitting content into MDX, wiring navigation and opening PRs via Codex Cloud. Teams relied on the Codex CLI, IDE extension, Cloud and best-of-N features to run asynchronous tasks, generate UI mockups, write evals, refactor agent architectures (single vs multi-agent), and produce Mermaid diagrams for booth walkthroughs. This matters for the AI/ML community because it demonstrates how code-centric LLM tooling can materially change engineering workflows — enabling rapid prototyping, parallel multi-tasking, and end-to-end app generation from protocol implementations to frontends, SDK bug fixes and documentation. The technical takeaway is that modern developer assistants can handle low-level protocol work (VISCA), full-stack scaffolding, and iterative design loops, but teams still used code review, evals and Guardrails SDKs to catch bugs and safety issues. The result is a clear productivity multiplier for ML teams, paired with a growing need for rigorous validation, testing and guardrails when automating critical engineering tasks.