Learning Library

← Back to Library

Comet Redefines AI Agents with UI

Key Points

  • The speaker has been inundated with AI agent pitches but found none truly impactful until discovering Comet, whose effectiveness stems from its superior user interface rather than raw AI capability.
  • Unlike other tools such as Zapier or n8n that require heavy effort to define and maintain specific workflows, Comet aims to function as a general‑purpose assistant that automatically handles tasks without the user needing to manage its inner workings.
  • Previous attempts at general‑purpose agents, like OpenAI’s Operator and Gemini’s Project Mariner, suffered from clunky, slow, and unintuitive UI designs that hampered adoption, highlighting the critical role of UI in agent success.
  • Perplexity’s Comet succeeds by making the assistant “disappear” – it executes work silently in the background, delivering results without exposing the user to the complexities or cost of building and controlling the agent.
  • The speaker emphasizes that a well‑designed UI can turn an AI agent from a novelty into a practical, day‑to‑day productivity tool, and Comet exemplifies this shift.

Full Transcript

# Comet Redefines AI Agents with UI **Source:** [https://www.youtube.com/watch?v=rkQhpiLn8EI](https://www.youtube.com/watch?v=rkQhpiLn8EI) **Duration:** 00:11:49 ## Summary - The speaker has been inundated with AI agent pitches but found none truly impactful until discovering Comet, whose effectiveness stems from its superior user interface rather than raw AI capability. - Unlike other tools such as Zapier or n8n that require heavy effort to define and maintain specific workflows, Comet aims to function as a general‑purpose assistant that automatically handles tasks without the user needing to manage its inner workings. - Previous attempts at general‑purpose agents, like OpenAI’s Operator and Gemini’s Project Mariner, suffered from clunky, slow, and unintuitive UI designs that hampered adoption, highlighting the critical role of UI in agent success. - Perplexity’s Comet succeeds by making the assistant “disappear” – it executes work silently in the background, delivering results without exposing the user to the complexities or cost of building and controlling the agent. - The speaker emphasizes that a well‑designed UI can turn an AI agent from a novelty into a practical, day‑to‑day productivity tool, and Comet exemplifies this shift. ## Sections - [00:00:00](https://www.youtube.com/watch?v=rkQhpiLn8EI&t=0s) **UI‑Driven AI Agent Success** - The speaker argues that, unlike other AI agents requiring heavy setup, Comet’s intuitive interface lets the assistant operate autonomously and seamlessly, making it the only agent that genuinely improves their workday. - [00:03:24](https://www.youtube.com/watch?v=rkQhpiLn8EI&t=204s) **Seamless Sidebar Automation with Comet** - The speaker highlights how Comet’s sidebar lets them edit approvals, send emails, manage calendars and LinkedIn, and discover local restaurants instantly—emphasizing speed, trust, and an unobtrusive, efficient UI. - [00:06:27](https://www.youtube.com/watch?v=rkQhpiLn8EI&t=387s) **Comet AI Identifies TikTok Layoffs** - The speaker praises Comet’s coherent analysis of Amazon’s strategy and demonstrates its capability to automatically pull LinkedIn profiles of TikTok Shop product managers and engineers possibly impacted by recent layoffs, all while integrating with calendar and email tools. - [00:10:53](https://www.youtube.com/watch?v=rkQhpiLn8EI&t=653s) **Rapid Rise of General-Purpose Agent** - The speaker highlights an unexpectedly fast, efficient AI agent that seamlessly handles diverse tasks across multiple platforms, suggesting it may outpace the anticipated rollout timeline. ## Full Transcript
0:00My inbox is littered with people 0:02pitching me AI agents. If I go online, I 0:04see nothing but AI agents. It's the year 0:06of the AI agent, and I have yet to find 0:09an AI agent that really made a 0:11difference in my workday until today. 0:14Comet is that AI agent. And the reason 0:17why is not AI, it's UI, it's user 0:21interface. And I'm going to get into why 0:22that is and kind of show off Comet a 0:24little bit. at the end of the day is an 0:26agent that works because the team at 0:29Perplexity has figured out that my goal 0:32is to actually let the assistant do the 0:34work and disappear. It is not to see 0:36everything the assistant does. It is not 0:38to get buried in the cost of building 0:40the assistant. And so many agents right 0:43now bury you in the cost of building the 0:46assistant or controlling it directly. 0:48Let me give you a few examples. Zapier 0:50and NA very different generation. Zapier 0:52is a much older company. and it ends 0:54much newer. Both of them lean you into 0:56building an agent. You have to define 0:58exactly what you want it to do. You have 1:00to invest really, really heavily in 1:03getting it to work right. N8 is soaring 1:05in popularity because it is better at 1:07that. It does offer value. Don't hear me 1:09wrong, but it is not a general purpose 1:12agent. It's not close to a general 1:14purpose agent. And if you try to build a 1:15general purpose agent in NA, you're 1:17probably going to be disappointed. It is 1:19useful for specific agentic tasks. If 1:22you want to do a specific job that 1:24involves extracting text from a document 1:26and putting it into a spreadsheet, for 1:27example, Naiden will do fine at that. 1:29You can probably get Zapier to do fine 1:30at that. There are a lot of other agent 1:32startups that will do that, too. But 1:34Comet is different. Comet by Perplexity 1:36is really a generalurpose agent 1:39assistant. The closest comparison we've 1:41had previously has been Operator from 1:43OpenAI and Project Mariner, which most 1:46people don't know about from Gemini. 1:48Look, I wanted to love Operator. I 1:50really did. operator. The vision was 1:53great. Chad GPT on the web, the power of 1:56Chad GPT decisioning across the web. The 1:58execution and the UI were really, really 2:01bad. They just were, I'm not going to 2:02sugarcoat it. It is awkward to have this 2:06tiny little browser that looks like a 2:08toy-sized browser inside a chat window. 2:10It's slow. Like, I don't care what the 2:14actual clock time says on these Asian 2:16things. I think it starts the clock late 2:17or something because I swear it takes 2:19longer than it says it takes. I've had 2:21instances where operator says it took 2:23eight minutes and it was like 20 minutes 2:25of clock time. And so if you think about 2:27it, the vision for operator was great, 2:30but what doomed it was the UI. It was 2:33the UI. And what Perplexity gets right 2:35is the UI. That's why Comet shines. 2:39Comet shines because the fundamental 2:41inside of the Perplexity team is that 2:43the assistant should disappear. They 2:45should just go do work for you. I was on 2:48a different website today. I was 2:50actually looking at Substack cuz of 2:52course I was and I was just like crap I 2:54have a meeting I have to reschedule it 2:56and so I just worked with Perplexity to 3:00get the meeting rescheduled. I talked to 3:01my common assistant and I said hey this 3:04is the meeting approximately here is 3:06where I want it moved. There's a block 3:07of meetings here. Suggest a time. It 3:09went it found the meeting. It suggested 3:10the time. It actually gave me a 3:12recommendation for a better time within 3:13that block that would fit well with my 3:15goal of like clean working blocks. And 3:17then it drafted up um a calendar change. 3:21I made a minor edit and approved it. And 3:24then it said, "Hey, just as a best 3:25practice, you should send an email to 3:27this person, too." And I said, "Okay, 3:28great. I'll send an email." I drafted up 3:30the email and I made another minor tweak 3:32and approved it and sent it all from the 3:34sidebar while I was doing other things. 3:36I didn't have to go to Gmail. I didn't 3:37have to go to Gcal. It was super easy. 3:39LinkedIn. I was able to go through and 3:41look at my pending LinkedIn invites 3:43sitting in the sidebar chatting with 3:46Comet. Super super easy. What I'm saying 3:48is that the idea that we have to see the 3:51agent is probably a legacy of the idea 3:53that the agent is untrustworthy. And if 3:56the agent can hook into data and it just 3:58works, we don't need to supervise it as 4:00closely and we kind of don't want to. We 4:02want it to be fast and trustworthy and 4:04dependable. And I will say comet gets 4:06this stuff done quickly. It gets it done 4:08really well. I am a fan, as you might 4:10expect, of Indonesian cuisine. It found 4:13for me an Indonesian restaurant that I 4:16had not run across in my neighborhood. 4:20Well, not quite my neighborhood, two or 4:21three neighborhoods over, but still, it 4:23found something in a special interest of 4:25mine in my city that I wouldn't have 4:27been able to find otherwise. So, I'm 4:29getting real value on day one, and it's 4:32been really easy. I also think when 4:35Comet needs to interact with your 4:37browser and starts to take over, the way 4:39they've done that UI is really good. 4:41What Comet does is it actually goes to 4:43your main browser and glows it up blue 4:45like you're like looking through a Star 4:47Trek portal or something and then it 4:48like does something quickly in the 4:50browser and then like it takes the blue 4:52away. It's a very intuitive way of 4:53saying, "Hey, Comet is driving now. You 4:55always have a chance to stop it." It's 4:57just really smoothly implemented. You're 4:59probably tired of hearing me talk about 5:02Comet. So, let's very quickly show you 5:04Comet. Okay, here we are. Tik Tok shop 5:08is slowing down. That is I literally put 5:11that into the address bar here. It's 5:12triggers a perplexity search. It talks 5:14about the impact of how Tik Tok shop is 5:16slowing down. I have a thesis that I 5:19want to explore with assistant and we're 5:21just going to watch Assistant work. 5:23Okay, I have a thesis 5:26that Tik Tok shop slowing down affects 5:30social commerce more broadly and may 5:34impact 5:35Amazon's outlook. 5:38Please look at Amazon's 5:42recent public info 5:45filings, etc. Get a sense of how the 5:48company is looking at social commerce. 5:52And you can see it's just working away 5:53there. Now if I want to continue, I can 5:55also continue here. I can say 5:59has Tik Tok experienced layoffs around 6:03Tik Tok shop right this is happening 6:05simultaneously. So Amazon's produ 6:07pursuing an agnostic social commerce 6:09approach explicitly response to Tik Tok. 6:14uh similar integrations with Instagram, 6:16Snap, Pinterest, and YouTube Shorts. 6:20Growing 6:22a bunch of other initiatives to keep up 6:24with creatives. 6:29So, I'm getting a sense of how Amazon's 6:31approaching this, 6:34and I really appreciate this is actually 6:36a more coherent take than I typically 6:38see from AI at this point. 6:43It's it's not just bland like Amazon is 6:47not ignoring tech stock slowdown. It's 6:48hedging risk and investing further. That 6:50seems to be supported by the comment up 6:52there. And it's also more actionable and 6:55on point than I've typically seen. Uh 6:57and yes, this has affected Tik Tok shop. 7:00So, it seems like we have a Tik Tok shop 7:02impact. All right, let's just do 7:03something fun. Can you find the uh top 7:08profiles on LinkedIn for Tik Tok shop 7:12affiliated 7:13PMs and engineers 7:16who may have been affected by layoffs 7:20and it's going to go and like look at my 7:22LinkedIn. So what what I'm trying to 7:24give you a sense of is how this can look 7:28across multiple things. It plugs into 7:30LinkedIn. It plugs into your calendar. 7:32It plugs into your Gmail. You can do 7:34multiple things at once. This is going 7:36to keep running even if I change where 7:38I'm going up here in the address bar 7:43and then it starts to find them and I 7:45don't know like I'm not going to go and 7:46click on their links. I don't want to 7:48violate their privacy. But like it gives 7:50me both a way to approach it and some 7:52example profiles. Overall it's pretty 7:54good. 7:56So there you go. That's Comet. I have 7:59been really pleased with it. I think 8:01that Comet is the first agentic browser 8:05that actually lives up to the name. And 8:07I think whether it's worth $200 a month 8:10essentially requires you to add up those 8:125, 8, 10, 15 minute increments that it's 8:14going to be saving you and be 8:16disciplined about it, measure the value 8:18of your time and say, is this worth $200 8:21a month in time savings to me? And 8:23that's a new way of valuing software. 8:25But I think that's where we're at with 8:26cognitive uh intelligence baked into 8:29software at this point. It gives us a 8:31new valuation paradigm for software that 8:35I'm still getting used to myself. I'm 8:37not used to paying $200 a month for any 8:39kind of software, let alone for a 8:40browser. Old Nate is crying. I don't 8:43know why I'm doing this. But it makes 8:45sense when you think about the time 8:46savings. If this thing saves me 10 15 8:49hours a month, it's obviously worth the 8:51$200. And if you add up all those 10 15 8:54minute increments saved through the day, 8:56I can see the case for 10 hours a month 8:58in savings and that would make it worth 9:00it. So I think my challenge to you if 9:03you want to start to explore good agents 9:05is to ask yourself to impose the kinds 9:08of tests you would actually use them 9:10for. Be really rigorously honest with 9:12whether they work or not. Like I put 9:14Comet through eight different workflow 9:15tests this morning just to see how good 9:18it is before making this video. I was 9:20not easily impressed. I wanted to 9:22examine it and I wanted to examine it 9:24versus operator because operator is a 9:26tool I know pretty well and been frankly 9:27disappointed in. If you want to try 9:31Comet, I would suggest signing up for it 9:34for just a month and putting it through 9:36workflow test in your own environment 9:38with the tools you use and see if you 9:40get value and measure the value. Measure 9:42the time savings. If you're going to 9:44make an investment like this, make sure 9:46you make it worth it. Because if it is 9:47worth it, if it saves you more than 200 9:49bucks a month, it's going to be probably 9:52a multiple of that. Like, if it's saving 9:54you 10 hours and you value your time at 9:56$35 an hour, you're getting close to 2x 9:59ROI every single month. If you value 10:01your time more highly, you get more. If 10:03you save more time, you get more. You 10:04see where the math goes. And that 10:06doesn't count it helping you make better 10:08decisions. And because it can look at 10:10the browser natively, it can help you 10:12make better decisions in a way that like 10:14screenshotting doesn't really do. And so 10:16there's another sort of hard to measure 10:18aspect where there's something native 10:20about the collaboration between the 10:21assistant and the browser that you can't 10:23get any other place cleanly. And I think 10:25that's another piece of value that 10:27perplexity is positioned to capture. Now 10:29the key if you're perplexity is that you 10:31are making an OS, right? This is where 10:33all of these AI players are going. We 10:35live on the web so much that if you 10:37become the browser, the dominant browser 10:39of choice, you become the OS for AI. And 10:42that's where they're going. Time will 10:44tell if this actually pays off. If 10:46Google launches Mariner more 10:47aggressively because Google obviously 10:49has a stake in the ground with Chrome. 10:51Maybe Open AI is going to lean away from 10:53voice a little bit and lean farther into 10:55operator. Who knows? One thing we know 10:57about AI is that it's going to change 10:58frequently. But for now, what I know is 11:01that I actually saw a general purpose 11:03agent that work today. And I'm still 11:05amazed. I never expected to see a 11:08general purpose agent that was this 11:09fast, this efficient, this effective 11:12across a wide range of tasks, flight 11:14reservations, restaurants, looking at 11:17LinkedIn tasks, looking at Gmail tasks, 11:19looking at calendar tasks, looking 11:20across multiple different sources to 11:22accomplish workflows. It's doing a lot. 11:24And I can see that the team has built a 11:26data structure that's going to enable 11:27them to do even more. To me, this feels 11:30ahead of schedule. Yes, I knew agents 11:32were coming, obviously. Yes, I knew 11:34general purpose agents are coming. This 11:36fast and this connected feels three or 11:39four months early. So, it's great to see 11:41it. And we'll see sort of where the 11:43Perplexian Comet team takes us next. Let 11:45me know if you get comment what you 11:46think of