At NVIDIA, engineering and research teams are adopting Codex as their default tool for complex engineering and autonomous machine learning experiments. Built on GPT-5.5 and running in production on NVIDIA GB200 and GB300 infrastructure, Codex handles longer, more autonomous sessions that go beyond execution to surface issues and ideas not present in the original prompt. According to Dennis Hannusch, a senior software engineer on NVIDIA's coding agents team, the system requires much less handholding and maintains high accuracy across multiple compactions during long sessions. Hannusch used Codex to evolve an internal platform from an MVP into a production-ready system and built a private podcast recording app in hours using the Codex desktop app with computer interaction, which also tested the video and audio features autonomously. For NVIDIA's research teams, Codex automates the research loop by identifying areas from large corpuses of papers, writing training scripts, and executing machine learning workloads on remote machines via SSH support. AI researcher Shaunak Joshi noted that GPT-5.5 acts as a creative partner that traces snippets of evidence and suggests knowledge graphs to visualize concept connections. Overall, Codex is accelerating work at NVIDIA by taking ideas from concept to execution and testing within a single workflow.
How NVIDIA engineers and researchers build with Codex
NVIDIA engineers and researchers are utilizing Codex, powered by GPT-5.5 and running on NVIDIA GB200 and GB300 infrastructure, as a default tool for complex engineering and autonomous machine learning experiments. The tool enables longer, more autonomous working sessions with minimal oversight, helping teams move from concept to execution and testing within a single workflow. Staff members note improvements in autonomy, context retention, and tactical tool selection compared to earlier models.
Why it matters
The deployment of GPT-5.5-powered Codex across NVIDIA demonstrates how advanced, highly autonomous AI coding agents can drastically accelerate internal engineering workflows and automate complex research loops. This points to a future where software development and experimental research scale rapidly with minimal manual intervention.