OpenAI Codex study finds agents overtaking chat in real work

This paper is one of the first large‑scale, data‑rich looks at how agentic AI is actually being used inside a frontier lab and across external organisations. The headline is stark: within OpenAI, chat is basically dead for knowledge work. Employees and many enterprise users now route almost all serious tasks through Codex-style agents that execute long‑running, multi‑step workflows. That’s a qualitative shift in how software engineering, analysis and even legal work are organised.

For the race to AGI, that matters more than another benchmark win. It suggests that agentic systems are already changing labour composition and workflow design in ways economists used to model as speculative. Once organisations learn how to manage fleets of agents – set goals, verify outputs, orchestrate parallelism – they create organisational “docking ports” for progressively more powerful models. That lowers the friction to adopting each new generation of capabilities, compounding their impact. It also surfaces new failure modes, as autonomy pushes beyond what current governance and verification tools can safely oversee, which will be the next bottleneck.