Comet Assistant will now ask before making important decisions: Aravind Srinivas

Aravind Srinivas, the CEO of Perplexity, has announced exciting upgrades to the company’s Comet Assistant, focusing on enhanced user control and improved safety for online tasks. During a recent update shared on Saturday, he emphasized how these changes empower users to better understand and manage the assistant's actions. The latest improvements provide users with clearer insights into when Comet is active, what tasks it is performing, and how it operates on their behalf. In a comprehensive blog post, the company detailed the revamped functionalities, which now allow users to see precisely when Comet is browsing, where it clicks or scrolls, and the specific steps it takes to find answers. One of the standout features is the introduction of a sidecar panel that enables users to track Comet's reasoning and observe its actions in real time. This transparency is designed to enhance user confidence and trust in the assistant's capabilities. Additionally, Perplexity has refined the way Comet handles tasks, introducing a prompt that asks users for permission before taking control. When a user enters a query, Comet can now request whether it should proceed autonomously. Users have the flexibility to choose between manual browsing, allowing one-time assistance, or granting ongoing permissions for automatic help. The company states that this design keeps users in command while still leveraging the assistant's speed and efficiency. Particularly for higher-stakes tasks, such as logging into accounts or making purchases, Comet will pause and seek explicit consent before proceeding, emulating the judgment of a responsible human assistant. These enhancements form part of a larger initiative by Perplexity to develop reliable and transparent AI tools within its ecosystem, which includes offerings like Deep Research, Labs, Email Assistant, and Background Assistants. By prioritizing visibility, user agency, and careful decision-making, Perplexity aims to shift the focus from merely querying what the assistant can do to actively deciding what users want it to accomplish.