The AI War Is No Longer About Better Answers. It’s About Getting Permission

Until recently, there was a fairly comfortable rule in our relationship with artificial intelligence: we asked, and the machine answered.

It could be wrong, invent a citation, write a terrible poem or finish in thirty seconds what used to take two hours. But the boundary was easy to understand. Initiative still belonged to the person on this side of the screen.

On September 29, OpenAI introduced Dots: agents with their own cloud computers that can connect to thousands of applications and keep working toward a goal without being directed through every step. OpenAI says they can learn from feedback, keep several projects active and work persistently. The same day, the company added computer use to the Agents API and multi-agent capabilities to GPT-6.1 Sol.

Google has been moving toward the same destination from another direction. On September 23 it expanded Gemini Connected Apps with services including Airtable, Linear, monday.com, Adobe, Webflow, Squarespace and Zoho. Eight days earlier, it introduced Gemini 3.8 Live and Live Extended Thinking, models that can run tools and background tasks while the conversation continues.

Taken separately, these are product announcements. Taken together, they suggest something more consequential: the prompt is beginning to lose its place as the main interface for AI.

The prompt may turn out to be a phase

There was a time when using a computer meant learning commands. Then came windows, icons and the mouse. Commands never disappeared; they simply stopped being the front door for most people.

AI may be heading through a similar transition. Today it still sounds sophisticated to write: “Find these documents, compare them with last month’s versions, update the table and tell me if anything looks unusual.” But once an agent knows the project, can read the sources it has been allowed to read and remains active, an awkward question follows: why should it wait for someone to type the same prompt again?

OpenAI is already describing that behavior. A dot can notice a bug reported in Slack and start investigating it; it can update analyses as new data arrives; it can watch a project for changes and prepare material for approval. OpenAI even uses the phrase “proactive research” for background review of connected sources, although in that mode it limits tools to read-only access.

That changes the relationship with software. A chatbot waits. An agent watches a state, notices that something changed and starts working inside the boundaries it has been given.

The problem is no longer just how much it knows

We became used to comparing models with questions that fit neatly in a table: which one reasons better, which one codes better, which hallucinates less, which has more context. Those things still matter. But an AI that acts introduces a less glamorous and possibly more decisive variable: what we allow it to do without asking us first.

Google is surrounding Gemini with applications. OpenAI is giving agents computers, browsers and external connections. Both strategies need the same thing to become useful: access.

Email. Calendar. Files. Corporate tools. Internal systems. Later, in some cases, purchases, payments or devices.

Inside a company, this looks less like installing a chatbot and more like hiring someone. A person can be brilliant and still not receive, on Monday morning, the company bank credentials, permission to delete databases and authority to send contracts. First they observe. Then they prepare. Later they execute certain tasks. Over time, some controls disappear because trust has been earned.

Agents may compress that process dramatically. That is why OpenAI spends so much time on rules, permissions and approvals. Dots can distinguish between actions that may run, actions that need review and actions that are blocked. Some sensitive operations remain reserved for the user. That is not a safety ornament added after the product. It is part of the product architecture.

Before deciding for us, they will decide what deserves our attention

There is another consequence that is easier to miss. If a system knows our goals, continuously observes what is happening and can act, it no longer participates only after a human decision. It starts participating before the decision by selecting what reaches our attention.

An agent reads a hundred emails and puts three in front of us. It did not reply on our behalf, but it already performed an important cognitive operation: it decided which three appeared to deserve time. It reads twenty documents and flags two anomalies. It compares several options and shows one first.

That is an old problem from psychology wearing new clothes. Human beings do not decide over the full universe of possibilities. We decide over what we see, remember and have in front of us. Anchoring, availability and default effects do not require an AI to make the final decision. It only has to arrange the stage on which the decision will be made.

That is why the word assistant begins to feel too small. A traditional assistant receives assignments. A system that filters, prioritizes and acts persistently helps construct the problem the user will eventually see.

Autonomy will need geometry

It is revealing that the industry is moving in two directions at once. It wants agents that need fewer instructions and, at the same time, more precise mechanisms for knowing when those agents must stop and ask for permission.

OpenAI says Dots examine actions that could affect accounts or share information and compare them against instructions, custom rules and safety requirements. Some tasks can proceed; others require approval; some must be performed directly by a person. Google, meanwhile, is pushing Gemini into more applications and toward Live models that can run tools while keeping the conversation going.

The race, then, is not simply about maximizing autonomy as if it were one knob. Autonomy has to be designed.

What it can read. What it can prepare. What it can modify. What it can send. What it can buy. What it can delete. How long it may keep working. When it must ask. And who is responsible when it did exactly what it was allowed to do, but the result was still a bad idea.

OpenAI and Google are attacking the same problem from opposite sides

Google has an obvious advantage: it already lives inside Gmail, Docs, Search, Android, Maps and Workspace. Gemini can become an intelligent layer crossing services that millions of people already use every day.

OpenAI does not control that ecosystem. Its bet looks different: build agents that can cross applications from many providers, carry context among them and operate from their own cloud computers. OpenAI says Dots can connect to more than 4,000 apps through its plugin ecosystem.

One company is trying to bring AI to the places where the user already is. The other is trying to build an AI that can go almost anywhere the user works.

Both strategies eventually hit the same door: permission.

Maybe we are watching the wrong race

There will keep being comparisons between GPT-6, Gemini 3.8 and whatever model arrives next week. There will be benchmarks for math, coding, reasoning and voice. One model will sit at the top for a few days and another will replace it.

But September’s launches suggest that the competition is moving into a field that is harder to score: who can embed intelligence reliably inside everyday work, and how much autonomy users are willing to hand over.

The day this works well, we may write far fewer prompts. Not because we finally learned to write the perfect prompt, but because the machine learned when we would normally have written one.