Meta Tests Autonomous AI Agent Muse for Web Browsing and Digital Commerce

Meta Platforms is developing Muse, a consumer-facing artificial intelligence agent designed to execute multi-step real-world tasks through automated web browsing and third-party account integrations. Users interact with the software through customizable virtual personas—including a default animated avatar named "Marley"—to delegate administrative chores and e-commerce transactions, according to a hands-on report from The Verge (https://www.theverge.com/ai-artificial-intelligence/999526/meta-muse-ai-agent-hands-on-shopping).

In hands-on testing across digital productivity workflows, Muse connected directly to Google's Gmail to inspect communication history with a local contractor. After confirming that an estimated bid deadline had passed, the agent drafted an email follow-up and shortened the copy upon user request before the message was approved and sent.

The agent demonstrated functional e-commerce capabilities when connected to payment credentials via Stripe. Operating within virtual browser sessions, Muse located replacement components for consumer appliances, populated customer address fields, and executed checkouts for consumables such as espresso water filters and descaling powder. During product sourcing, the model performed price-efficiency calculations between powdered and liquid formulations before completing orders.

However, the software encountered friction when navigating anti-bot protections on major retail platforms. When directed to search for outdoor camping supplies, Muse was blocked by Amazon for violating site terms of service governing automated software agents. The agent subsequently routed its search to alternative merchants, finding items across Walmart and direct-to-consumer websites such as OXO.

Muse also handled local service coordination by extracting user-provided photos, street addresses, and phone numbers to complete online intake forms. When tasked with arranging yard debris removal, the agent navigated regional hauler portals and submitted quote requests, yielding vendor follow-ups via text message with minimal direct user interaction.

Despite its web navigation capabilities, Muse remains limited when interacting with legacy telecommunications systems. In a test seeking to downgrade a home security subscription, the agent logged into the user's portal to retrieve account details but could not place the phone call mandated by the provider to finalize the change. Voice-calling features for the agent are reportedly in development, but current workflows still require manual human intervention for phone-based customer service.

Deploying Muse requires substantial user data access. Full feature execution during testing involved providing the agent with names, home addresses, Stripe payment authorization, residential photos, and third-party login credentials. Meta states that user authentication tokens and sensitive information are stored securely and remain inaccessible to the front-end agent interface.