Charles Gabriel already said where the human sits in an automated loop
On Ep13, 95% of enterprise AI projects fail, Charles Gabriel put it plainly: "you got to have human in the loop right now… giving AI context for a task, that has to come from the human… once it has enough context for a certain task, it could potentially, you know, just continue to automate that task. But how you query, how you prompt, the type of context you provide, that's all within the control of the human."
What it actually means
OpenAI has now measured the thing Charles described. Inside the most agent-native research team there is, planning is a sliver of what the agents produce, people still set the priorities and judge the results, and more than half of the long tasks that succeeded needed someone to step in. Charles's split holds at 3.1 agent-days per human day: the agent runs the task; the human supplies the context and the judgement. If that is where the human sits at OpenAI, it is where the human sits in your workflow too. Design the job around the steer, not the task.
The agent runs the task. The human supplies the context and the judgement.
Source: Brew Bytes Ep13 — Charles Gabriel · Tuesday 8 September 2026