Can an AI browser actually run your errands? We tested the real number

The exact clicks

1. Go to perplexity.ai/comet and click "Get Comet" — it's a small, fast download (about 13MB), free, no account required to start.
2. Open the downloaded file and install it, same as any other browser (Chrome, Edge).
3. On first launch, sign in with a free Perplexity account, or create one.
4. Comet will offer to import your existing browser's history, bookmarks and passwords — this step is optional, the browser works without it.
5. Open the Comet Assistant (the sidebar panel — its exact icon/label can shift between versions, so if it's not obvious, check Comet's own onboarding tour, which points it out).
6. Type a real task in plain English — e.g. "find a flight from my city to London next weekend and show me the three cheapest options."
7. Watch it work: Comet shows each step it takes as it browses, not just a final answer — you can stop or correct it mid-task.
8. Start with a short, one-destination task before trying anything with multiple steps chained together — that's exactly where it's most likely to break (see Limits).

What changes

• The real number, from an independent test across 136 real websites: the strongest AI browser agent completed 61.3% of 300 everyday tasks; most agents scored closer to 30%.
• A human given the identical list of tasks completes nearly all of them — the gap between that and even the best agent is the part a product demo never shows.
• Comet's free tier already includes agentic browsing (form-filling, shopping, multi-step workflows) — this isn't a paywalled preview feature.

What it will not do

• It does not reliably chain many steps together — a wrong turn early in a long task compounds, and everything after it goes wrong too. Shorter, single-purpose tasks are where it's actually reliable today.
• A paid "background assistant" tier exists for running tasks without watching them live — the free tier expects you to stay present and check in.
• This episode names one specific browser (Comet) and one specific benchmark, both current as of publishing — re-check both if this episode is more than a few months old; this category is moving fast.
• Not a security review: agentic browsers as a category have documented risks around a malicious page hijacking an agent's actions — this episode doesn't cover that side, only whether the everyday-task claim holds up.