From questions to delegation, the era of always-on agents

For three years we used AI the same way: one question, one answer. In 2026 that pattern is breaking, and I think it’s the most important change of the year, more than any leaderboard point.

Three signs of the same shift

The answer becomes an interface. With GPT-6, in October 2026, ChatGPT introduced Intelligent UI: ask “how do I split the holiday expenses?” and instead of text it builds a small tool on the spot with fields to fill in. Charts, maps, forms, calculators. The answer is no longer something to read, it’s something to use.

The assistant works while you’re away. “Always-on” agents (OpenAI calls them Dots and brought them to the Pro plan) follow a task for days: they watch the inbox, update a document, alert you when something happens. Scheduled tasks, available in all major assistants, are the simple version.

Agents organise among themselves. Anthropic is testing a mode to manage sub-agents from the app; OpenAI described experiments with thousands of agents working for over 36 hours; Meta presented Muse Wonder, built precisely for agents; Microsoft launched a model, Decision-1, whose job is to route and control other models’ work. The single model becomes an organisation.

From “asking” to “delegating”

The shift is subtle but huge. When you ask, you check each answer before using it. When you delegate, the work goes on without you, and you see the result at the end (or an exception when something goes wrong).

It’s the same leap a professional makes going from working alone to having an assistant. And as with a human assistant, the questions change:

what can I hand over, and what not?
which data and tools should it access?
how will I notice if it makes a mistake?
who answers if the mistake reaches the client?

The bottleneck is trust

Today’s agents are good but not infallible. On Arena’s Agent leaderboard, built from millions of real sessions, even the best have margins far from perfection; and agent security (prompt injection above all) is an open problem.

That’s why I think over the next two years the difference between who gets value from agents and who doesn’t won’t depend on the chosen model, but on three very non-technological things:

Written processes. An agent can’t follow a procedure that exists only in your head. Organisations that have documented how they work will delegate much more.
Tidy data. Consistent folders, clear names, one version of each document. The agent is as fast as the mess allows.
Thought-through permissions. Who can read, who can write, who can send or pay. The same rules you’d give a new hire, applied to software.

The one-person company?

There’s a lot of talk about the “one-person company”: one person and a swarm of agents doing a company’s work. There are real, interesting examples, especially in software and content. But for most businesses I think reality will be more modest and more useful: the same team doing more and better, because repetitive tasks run in the background.

In a professional firm it means the invoice arrives, is read, proposed for logging and queued for review before anyone opens the inbox. The professional starts the day from the exceptions, not the pile.

What I’d do now

Start with scheduled tasks: they’re low-risk agents and teach you to delegate.
Write down your procedures, even briefly: they’ll become Skills and instructions for agents.
Connect tools gradually, read-only when enough.
Keep a human on irreversible actions, and design the check to be realistic.

In one line

The near future isn’t an AI that answers better, but an AI that works while you do something else. The question isn’t “which model?” but “what am I ready to delegate, and under which rules?”.

Sources

Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.