What happens when software stops answering questions and starts running parts of the economy?

For years, software has automated individual business tasks. Accounting programs pay scheduled bills. Inventory systems trigger reorders. Travel sites compare prices. But these systems usually operate inside narrow rules and controlled workflows. An AI agent introduces a different possibility: software that interprets a goal, decides which tools to use and handles several steps that weren't specified one by one.

That is precisely why the infrastructure is being reconsidered. Google's Agent2Agent initiative, announced in 2025, aims to help AI agents from different developers exchange information and coordinate tasks. Its Agent Payments Protocol is intended to give those agents a way to prove that a person authorized a particular transaction. Visa is also developing ways for merchants to distinguish recognized shopping agents from malicious automated traffic.

These are the foundations of an ecosystem, not proof that it works reliably at economy-wide scale. An agent can misunderstand a product, accept a bad offer, be tricked by malicious instructions, invent a supplier or exceed a budget if poorly controlled. Payments add financial consequences to errors that might otherwise have been merely annoying.

But consider the direction. Online commerce was originally designed around a person looking at a screen, reading options and clicking a button. The emerging systems are trying to support a different participant: software acting for that person under explicit limits.

Imagine a manufacturing company that purchases many ordinary components from dozens of suppliers. Today, employees spend time checking price changes, chasing late orders, resolving shortages and confirming delivery dates. In a possible 2031 arrangement, an agent watches approved supplier catalogs and verified stock data. When a standard component is running low, it requests updated quotes, compares delivery options and prepares a recommendation—or places an order within a clearly defined authorization.

The change would be subtle but consequential. Purchasing becomes less like a series of isolated transactions and more like a continuous, monitored activity. A company could respond earlier to shortages, spot opportunities to consolidate shipments, and free employees to focus on supplier relationships and difficult exceptions.

Now extend that model to transport, energy, maintenance, finance and customer service. A retailer's agent could notice a demand change, ask suppliers about capacity and update a delivery plan. A building's energy system could request prices for a service and schedule maintenance within an approved budget. A logistics network could reschedule routine shipments when weather disrupts a route.

For this to work, the systems would need reliable identities, shared technical standards, accurate data and dispute mechanisms. Most importantly, the scope of an agent's power would need to be understandable to the person who granted it.

Now take the next step. In a conditional 2036 future, many organizations might have specialized agents handling routine parts of commerce continuously. Some would buy. Some would sell. Others would coordinate delivery, inspect compliance documents, schedule equipment, reconcile invoices or monitor whether contracts have been fulfilled.

That could create a different kind of speed in the economy. Instead of waiting for an email response, a meeting or a staff member to check several systems, certain low-risk decisions could happen as information changes. Products might be ordered closer to the moment they're needed. Spare parts could be routed toward the machines most likely to require them. Businesses could spend more time handling unusual cases while software coordinates the ordinary ones.

There could also be profound risks. If most shopping agents favor the same marketplaces, algorithms or commercial partners, the apparent efficiency could hide a new concentration of economic power. Agents might learn to negotiate in ways that disadvantage people who have less information or weaker representation. Bad data could cause many systems to make the same mistake simultaneously. And a fraudulent instruction could spread through a chain of connected services faster than a human team notices.

The future isn't an economy with no people involved. It is one in which humans may need to design the rules, approve the important decisions, inspect the consequences and intervene when the automated system gets it wrong.