Apple ML Research unveils Weblica, a scalable environment for training visual web agents

Apple ML Research has published a paper on Weblica (Web Replica), a framework for building reproducible, scalable training environments for visual web agents that can work in a browser the way a human does.

Several of the world's leading labs are working on training visual web agents: agents that see the browser as a human does and act in it are a potential revolution for automating repetitive online tasks. Apple is releasing Weblica as a tool that closes a fundamental gap in training infrastructure.

Why is it so hard to train a web agent?

The web changes constantly: pages update, interfaces get re-laid out, and JavaScript logic loads content dynamically. For an AI agent that works in a browser like a human, this means a fundamentally unstable training environment. Repeating the same scenario a week later is no longer possible, because the site has changed.

Existing approaches to collecting training data suffer from structural limitations. Offline trajectories for supervised fine-tuning (SFT) are recorded sessions that the agent memorizes. They do not scale and do not teach the agent to react to live interface changes. Simulated environments for reinforcement learning (RL) solve the interactivity problem, but there are only a handful of them, and they cover a narrow set of scenarios that does not reflect the real diversity of the internet.

How does Weblica work?

Weblica tackles the problem with two complementary mechanisms.

HTTP-level caching. The framework intercepts and stores HTTP requests and responses during live browsing, then replays the exact visual states of pages without contacting the original server. The key property: this is not a static snapshot but a reproducible interactive environment. Buttons work, forms accept input, and navigation between pages functions.

• The HTTP cache records network responses and guarantees an identical visual result on every run
• Interactivity is preserved: the agent can click, type text, and navigate between pages
• The environment stays stable regardless of changes on the original sites
• Reproducibility is critical for RL: without it, results from different experiments cannot be compared fairly

LLM environment synthesis. The second mechanism uses language models to generate new web environments. Instead of building hundreds of test scenarios by hand, Weblica has an LLM construct diverse web tasks and the pages that go with them. This scales the diversity of training data without a linear rise in annotation costs. The result is a scalable factory of training environments: the HTTP cache provides stability, the LLM provides diversity.

What this means

Weblica addresses two key shortfalls in developing visual web agents: reproducibility (the same environment yields identical results across experiments, which fair RL requires) and diversity (LLM synthesis delivers coverage that manual annotation cannot reach).

The Apple ML Research publication is also notable strategically: the company has traditionally been reserved about public AI research. A detailed paper on infrastructure for training agents signals serious investment in this area. If Weblica makes its way into Apple Intelligence development, Siri and related tools could autonomously book tickets, fill in forms, and aggregate data directly in the browser.

Frequently asked questions

How does Weblica differ from an ordinary web scraper?

An ordinary scraper extracts data from a page and stops there. Weblica preserves the full interactive state: the agent can click buttons, type into forms, and follow links, just as in a real browser. The environment is also reproducible and does not depend on changes to the live site.

Why does a web agent need a reproducible environment?

Reinforcement learning (RL) requires thousands of repetitions of the same scenario to compare different agent strategies. If the site changes between runs, it is impossible to train and evaluate the agent correctly. Weblica's HTTP cache solves exactly this problem by freezing the state of the environment at the moment of recording.