Codex, Behind the Harness — Dominik Kundel, OpenAI
Dominik Kundel from OpenAI presents Codex, a system for running AI models as agents—programs that can perform tasks autonomously.
When the model could run fast enough, the network became the bottleneck, so they switched to websockets for a constant connection instead of resending all data. The system manages which tools the model has access to by hiding rarely-used tools from the context window and only showing them when the model searches for them. All actions—such as editing files or running commands—are executed in sandboxes for security, and when someone requests something dangerous, an exception is triggered that a completely separate, protected AI agent reviews. The entire system is open-source and written in Rust.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
