OpenAI Wants Agents For Your Messy Life. ChatGPT Work Is The Trojan Horse. Good Luck With Websites From 1995.
OpenAI is developing ChatGPT Work, a modified version of its Codex coding tool, to bring agentic AI to non-engineers. The product aims to complete multistep projects autonomously rather than just answering questions. Ambrosino, a team lead, told TechCrunch the goal is building agents that can navigate the messy reality of outdated websites and personal tools.
This illustrates the principle of context transfer. The mechanism is domain adaptation: taking an agent trained on structured coding tasks and forcing it to operate in the unstructured mess of everyday digital life. The lesson for you is that agentic value comes not from better prompts but from an AI that can tolerate the chaos of legacy systems and incomplete instructions.
OpenAI is building ChatGPT Work based on its Codex tool. Ambrosino leads a team focused on expanding agentic AI beyond software engineering, per his TechCrunch interview.
- Open ChatGPT and ask it to plan a multistep personal task, like organizing a trip, breaking it into subtasks you can check off. Observe how it structures the work.
- Give it a messy real-world constraint, like a budget and three conflicting calendar invites, and ask it to reconcile them. Note where it succeeds and where it hallucinates.
- Try Claude Code or a similar coding agent on a simple file task to see what the engineering version of this looks like. The gap between the two experiences is exactly what ChatGPT Work is trying to close.