Before We Know What to Ask

GPT-6 Astra drawing on screen in OpenAI’s Introducing GPT-6 Astra.

Tony Stark says what he needs, and JARVIS gets to work.

In OpenAI’s demo, the presenter tells the model what they want. The model sees the screen, opens the app, and draws while they watch. Computer use lets the model take the steps itself, not just explain them. But being able to act does not tell it what is worth doing. It still needs us to point it toward something that matters.

Giving that direction is rarely as simple as issuing a command. We often discover what we want while trying to express it. An agent may be able to do almost anything, but we still have to work out what we want it to do.

A place to think

Suppose I start typing “Help me plan a weekend trip.” I pause, read it back, and realize that what I really want is a break from driving. I add “by train.” That small edit changes the trip the agent will plan. It also tells me something about what I wanted from the trip.

Putting the idea into words gives me a place to set the thought down while I decide whether it captures what I mean. I can change one part without having to hold the rest in my head.

Research on writing describes this as an ongoing process: we plan, put ideas into words, and revise our goals as we go.¹ Keeping those words visible gives us something to inspect and rearrange.²

What I say is only one part of what the agent might know. With permission, it may also have access to the tab I have open, the dates on my calendar, or the ticket I was considering. Those clues help it understand the situation, but they do not determine the goal on their own.

A shared, editable brief could bring that interpretation into view. As my thinking changes, I want to see what the agent now understands, including what it gathered from the context I allowed it to use. Does the trip still include a rental car? I could point to the part that feels wrong, keep the details that still fit, and continue from there, without needing to get every detail right at the start.

When the situation changes

Sometimes the missing part of a request comes from reflection. Sometimes it comes from the world changing around us. Giving an agent direction is not only about setting an initial goal. It is also about helping it notice when that direction may no longer fit.

Now suppose the trip is planned for tomorrow, with an afternoon outdoors. The forecast changes to rain. That gives me a new request: “Find an indoor alternative for tomorrow afternoon.”

I could not have included that detail in the original request. It became relevant later, when the weather changed.

An agent with access to my calendar and the forecast may have enough context to notice a possible conflict. What it still needs is a reason to look. A standing instruction to check the trip each morning could provide that direction. I would not have to anticipate every possible change, but I would still need to understand its suggestions. Showing the outdoor plan beside the updated forecast would make the suggestion easier to follow.

If I had already moved the outing indoors, I should be able to correct the agent right there and have that correction carry into its next check. If I still wanted to go outside, I could keep the plan. The suggestion would give me a concrete choice, with enough context to make it.

At Prelude, we want to help people give agents direction, see what they understand, and decide what should happen next. That means showing why a suggestion matters and giving people room to correct it before an agent acts.

By the time Tony Stark knows what to ask, JARVIS might already have the next mission mapped out, with two fingers of Laphroaig waiting.