Automating repetitive work at OpenAI with Codex

This post details how OpenAI is using Codex and the Runme notebook environment to automate repetitive engineering tasks, specifically focusing on running model evaluations. It describes a workflow where an agent (Codex) is given a goal in a Runme notebook, plans the execution, and documents the steps, commands, results, and decisions. The post highlights the use of WebMCP for client-side tool registration, enabling agents to interact with the Runme application directly from the browser without requiring a dedicated backend server. It also emphasizes the importance of collecting and curating context within the notebook to improve future agent runs and make practical knowledge shareable.