Apple ML Research Presents Weblica — A Scalable Environment for Training Visual Web Agents

Apple ML Research has published an article about Weblica (Web Replica) — a framework for creating reproducible and scalable training environments for visual web agents capable of working in a browser like a human.

Several leading laboratories around the world are working on the task of training visual web agents: agents that see the browser as a human and act within it represent a potential revolution in automating repetitive online tasks. Apple releases Weblica as a tool that fills a fundamental gap in training infrastructure.

Why is it so difficult to train a web agent?

The web is constantly changing: pages update, interfaces are redesigned, JavaScript logic dynamically loads content. For an AI agent working in a browser like a human, this means a fundamentally unstable training environment — it's impossible to repeat the same scenario a week later because the website 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 don't scale and don't teach the agent to respond 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 don't reflect the real diversity of the internet.

How does Weblica work?

Weblica solves the problem through two complementary mechanisms.

HTTP-level caching. The framework intercepts and saves HTTP requests and responses during live browsing, then reproduces exact visual page states without accessing the original server. Key property: this is not a static snapshot, but a reproducible interactive environment — buttons work, forms accept input, page transitions function.

• HTTP cache captures network responses and guarantees identical visual output on every run
• Interactivity is preserved — the agent can click, enter text, navigate between pages
• The environment is stable regardless of changes on the original websites
• Reproducibility is critical for RL: without it, it's impossible to fairly compare results from different experiments

LLM-synthesized environments. The second mechanism uses language models to generate new web environments. Instead of manually creating hundreds of test scenarios, Weblica delegates to the LLM the task of constructing diverse web tasks and corresponding pages. This allows scaling the diversity of training data without linear growth in annotation costs. As a result, Weblica creates a scalable factory of training environments: HTTP cache ensures stability, LLM ensures diversity.

What does this mean

Weblica addresses two key deficits in visual web agent development: reproducibility (the same environment produces identical results in different experiments, which is necessary for fair RL) and diversity (LLM synthesis creates coverage unachievable through manual annotation).

The Apple ML Research publication is notable also from a strategy perspective: the company is traditionally reserved about public AI research. The appearance of a detailed article about infrastructure for training agents signals serious investments in this direction. If Weblica enters Apple Intelligence development practices, Siri and related tools could autonomously book tickets, fill out forms, and aggregate data directly in the browser.

Frequently Asked Questions

How does Weblica differ from a regular web scraper?

A regular scraper extracts data from a page and stops there. Weblica preserves the complete interactive state: the agent can click buttons, enter text into forms, and follow links — just as in a real browser. At the same time, the environment is reproducible and doesn't 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 website changes between runs, it's impossible to correctly train and evaluate the agent. The HTTP cache in Weblica solves exactly this problem by fixing the environment state at the moment of recording.