I Was Wrong About AI Agents — This $200 Browser Actually Works

AI agents have been hyped for months, but most fail to deliver day-to-day value because they force users to supervise the process or spend too much effort configuring and controlling them. Comet, built by Perplexity, stands out less for “smarter AI” and more for a practical interface: an agentic browser experience designed to let the assistant do the work while the user stays mostly hands-off.

The core claim is that Comet’s advantage is UI—specifically, how it integrates with the browser and other everyday tools without turning the workflow into a slow, awkward control panel. The contrast is drawn with other agent products. Zapier and similar automation tools require heavy upfront definition of what the agent should do, which limits flexibility and increases setup cost. “N8” (as referenced in the transcript) is described as useful for narrow, agentic tasks like extracting text from documents into spreadsheets, but not close to a general-purpose agent. Even Operator from OpenAI is criticized for execution: the vision of a “ChatGPT on the web” browser is appealing, but the interface is portrayed as clunky and slow, with reported delays that feel far longer than the clock suggests.

Comet’s workflow is presented as the opposite: the assistant disappears into the task, then briefly signals when it needs to take over. In a scheduling example, Comet helps reschedule a meeting by finding the relevant time block, recommending a better slot based on the user’s preference for clean working blocks, drafting the calendar change, and then suggesting an email follow-up. The user makes minor edits and approves actions, while Comet handles the rest—without switching repeatedly to Gmail or Google Calendar. Similar convenience is described for LinkedIn, where pending invites can be reviewed from a sidebar.

The transcript also emphasizes trust and speed as design goals. The “agent should disappear” philosophy is framed as a response to a legacy assumption that agents are untrustworthy. If the agent can reliably hook into real data sources—calendar, email, browser content—then close supervision becomes less necessary. Comet is said to complete tasks quickly and to deliver tangible value immediately, including finding an Indonesian restaurant the user hadn’t encountered nearby.

A live demo reinforces the browser-native approach. Comet triggers a Perplexity search from the address bar, then iterates on a thesis about TikTok Shop slowing down and its implications for social commerce and Amazon’s outlook. It also continues the investigation by checking whether TikTok has experienced layoffs tied to TikTok Shop, and it surfaces a more actionable, coherent take than typical AI summaries. Another example shows LinkedIn integration: Comet searches for TikTok Shop–affiliated PMs and engineers who may have been affected by layoffs, then provides example profiles without forcing the user to click through every link.

Finally, the transcript reframes pricing as a time-savings calculation. At roughly $200 per month, the value proposition depends on whether Comet saves enough minutes across workflows to justify the cost—potentially 10–15 hours per month—plus the added benefit of making decisions using native browser context that screenshot-based tools can’t replicate. The speaker closes by calling Comet the first “agentic browser” that lives up to the promise of general-purpose automation, while acknowledging the market could shift quickly as major players push competing agent products.