OpenAI's own staff are switching from chatbots to parallel agents

OpenAI published a research note on how its own staff use AI agents, and the numbers describe a shift from chatting to delegating. The headline figure: by June 2026, the heaviest users (the 99th percentile) were running more than 60 hours of Codex agent work per day, spread across several agents working in parallel. That is only possible if you stop waiting for one model to type a reply and instead hand off many tasks at once, then check the results later.

The growth is uneven by function. Measured against November 2025, median agent use in Research rose 56 times, Customer Support 32 times, Engineering 27 times, and Legal 13 times. The sharpest change is who uses agents at all: non-developer individual users grew 137 times since August 2025, and non-developer organizations 189 times. OpenAI's read is that agents let teams without engineers get past work that used to require one.

Treat the data as what it is. This is OpenAI describing OpenAI, using its own Codex product, so the levels are not a neutral industry benchmark. But the direction matches what other tools report, and the company is specific about the mechanism: long tasks running in parallel rather than single prompts. You can read the full note on OpenAI's site.

Why it matters

If your team still treats AI as a faster chatbot, the useful question is which of your multi-step tasks could run as a delegated agent job instead. The teams pulling ahead are not asking better questions, they are running more work in parallel and reviewing the output.