Apple researchers taught AI to practise using apps without needing a real app to practise on

Key points

• Apple ML Research published a method in 2025 that trains AI agents to use software tools without requiring a live, working version of that software.
• The technique uses a large language model, the technology behind chatbots like ChatGPT and Claude, to act as a pretend environment that the learning agent practises inside.
• Current approaches to building API-calling agents depend on real databases full of realistic data, which is costly and slow to create.
• The new method needs only a written description of an API, short for application programming interface, the set of rules that lets one piece of software talk to another.

Imagine teaching someone to use a cash register without ever giving them a real till. You describe how it works, and a colleague plays the role of the machine. That is roughly what Apple ML Research has done for AI, except the colleague is also an AI.

Building an agent that can use software tools currently requires thousands of worked examples. Each is a recorded back-and-forth between the agent and an actual running application, complete with believable dummy data loaded in. Creating that at scale takes enormous effort.

Apple's approach sidesteps the problem entirely. Given nothing but an API specification, a document describing what a piece of software can do and how to ask it, the system conjures a simulated environment on the fly. A second large language model plays the role of the application, inventing plausible responses in real time. The learning agent practises against this stand-in, no real server involved.

This matters because the bottleneck today isn't the AI itself. It's the data pipeline. Researchers must build working test environments, populate them with believable dummy data, and run thousands of simulated tasks before an agent is ready for real users. That process limits how quickly new capable agents can be created.

Our coverage of Apple's virtual-user testing framework on 14 July showed the same lab thinking about the same problem from a different angle: how do you build and test agents without exposing real people first? This paper answers the earlier half of that question, the training side.

Developers would need only the documentation that software makers already publish, turning a months-long data-collection effort into something far faster.

What does this mean for people who use AI assistants?

For ordinary users, better-trained agents mean fewer frustrating mistakes when an AI tries to book a flight or manage a calendar on your behalf. If agents can be trained cheaply across many different apps, the range of tasks they handle reliably should grow.

The work is early-stage research, and Apple ML Research hasn't announced a product built on it. Real-world performance will need independent testing before anyone can say how well simulated training transfers to genuine applications. The underlying idea is a clean solution to a genuine, well-documented problem. What I'd watch is whether other labs adopt environment-free generation or whether Apple's implementation turns out to be harder to replicate than the paper suggests.