Lenny's Podcast: Product | Career | Growth
00:13:36 · Dan Shipper
[00:13:36] S1: That's a little part of your soul. I really think that's what it looked like was happening. And so I was very into personal agents and I have completely flipped and I really think that the model, for now, is going to be a super agent, like one agent for the entire company and you're starting to see this in some companies. So Shopify very famously has one, Ramp has one now and I think there's some really interesting reasons for that. I actually still think that the personal agent thing is coming, but what we found is there's all this hype with OpenClaw. Everyone's like, "I'm going to set it up. It's so cool," or whatever. And then everyone realizes it's like way too much work. This thing breaks all the time. I got to fumble around with it. I got to be able to SSH into my server and like blah, blah, blah. And most people, to do work at least, just don't want to spend that time, or can't. And the fundamental, underlying thing that drives that is whether it's OpenClaw or any other harness, in order for an AI agent to be useful right now, it really needs a human who cares about it. It really needs a human personal connection with someone who's watching what it does and makes sure that it's doing the right thing and that it's useful for people. And the minute you sever that connection, so the minute someone's like, "Ah, I don't want to maintain this dumb OpenClaw," is the minute the agent is not really that useful anymore. And that's why I think it has started to shift to a more one agent per company model because for now, the ideal is you basically set up a forward deployed engineer or someone with that profile who's responsible for making sure that that agent is working for the whole company and then maybe you have some little team agents. And I think as the models get better at being more independent, that will shift down and it'll be more likely that we'll have more personal agents because we won't have to fuck around with all the internals. But the model that I see working for us and for a lot of other companies, including the model companies, the model companies themselves are starting to see this, is when it comes to the like async agents, it's really a, you have one agent at the top that's like doing, sometimes it's everything.
00:16:02 · Dan Shipper, Lenny Rachitsky
A lot of times it's a particular job that you've decided that everyone in the company needs an agent for, like data requests. And then I think it starts at the top and then it starts to trickle down, where you may get more specialized agents and teams and all that kind of stuff. And the mechanism is agents need people who care about them. [00:16:24] S0: That is so interesting, that point about you need to garden your agent because there's context you have to keep adding to it. It breaks, as you said, and once it's just too much work, you're like, "Okay, forget this thing. I'm going to go back to Codex or Claude or something like that." [00:16:38] S1: Exactly. [00:16:39] S0: Okay, cool. So this is a cool opportunity. So the idea, so what you're predicting here is companies will have this super agent that everyone can talk to. As you said, Shopify's got River, I think it's called. What's the Ramp one called? [00:16:49] S1: I can't remember. [00:16:50] S0: Okay. It's probably got to fun name. Okay. So that's the prediction. Okay. [00:16:56] S1: That's the first prediction ... [00:16:57] S0: That's the first prediction. [00:16:59] S1: Is we will start with agents at the top that are more general and are used by more people in the company and then it will start to grow down, as people get more used to these use cases, they get more specialized and agents become less ... Fiddly. They just work better. [00:17:22] S0: And is this mostly going to be in Slack, do you predict?
00:17:22 · Lenny Rachitsky, Dan Shipper
[00:17:22] S0: And is this mostly going to be in Slack, do you predict? [00:17:25] S1: For work? Yeah, it seems to make sense. I think people love having the green bubbles on OpenClaw. Sorry, the blue bubbles on OpenClaw, if you can use it with your iPhone, but I think there's this little thing in people's heads where they really like to keep their personal and work agents separate. And I think there's a whole territory. Our COO, Brandon Gell, calls this computer errands. There's this whole territory of using personal agents for your computer errands. It's like order my groceries or whatever. And it's like there's so much of that that I think it's going to be huge for, but we focus mostly on the work stuff and I think that's going to happen mostly in Slack. [00:18:07] S0: Sweet. Go Slack. [00:18:09] S1: Do you want to talk about the other work surface? [00:18:09] S0: Absolutely. [00:18:12] S1: Like Codex, Cowork. Okay. This is the one that- [00:18:15] S0: Let's do it.
00:18:15 · Dan Shipper
[00:18:15] S1: I'm so excited about this one. I think it's the coolest thing. So basically what happened was Anthropic realized at some point that, with Claude Code, if you put an agent on your computer and it runs on your computer, it has access to everything that you have access to. It uses the terminal, so it has basically superpowered access to it. And not only that, these agents really understand how to use the terminal because there's so much content online about that and it created this super powerful coding paradigm, which is Anthropic was really doing it first. OpenAI, for a while, was, in my opinion, very, very behind on this and then, in my opinion, has surpassed them recently. It's really interesting, but they were very early on this. When people were still thinking about coding agents or coding models as being really pair programmers, they were among the first to be like, no, and do it successfully. There were people before them, like Devin, who, I think, had the big cloud environment and OpenAI tried this too, but the real adoption seems to have happened when you put it on your computer. So they figured that out and then I think they figured out, along with their community, that once you have a coding agent on your computer that can build anything, it's actually really good for any kind of work you want to do. And people started just hacking Claude Code, essentially, to do all of their work. Anthropic then built Cowork, which is a little bit of a nicer wrapping around Claude Code, but is fundamentally the same thing. And then I think OpenAI made a couple of different bets, but their main bet on a programming agent was the earlier versions of Codex were very technical and they were super smart, but they were a little bit autistic. It was a little hard to ... They didn't quite get what you meant. They got exactly what you said. And I think maybe like three or four months ago, around the time that they launched 5.3, they started to move in this direction of, "Oh no, we get it. This model is fast. It's really good for general purpose, knowledge work type tasks." And then they launched the Codex desktop app. And I think the Codex desktop app takes ... If you look at all the lessons that Anthropic learned, they went from Claude code to Cowork and you can see that in the tabs on the Anthropic desktop app UI.
00:20:45 · Dan Shipper
I think OpenAI was just like, "We see where this is going. Let's just skip to that." And so I think Codex right now, this is a horse race. They're going to have different positions, but I think Codex right now, it's my daily driver. I spend all my time in it, basically. I flip to Claude every once in a while, but I think they're getting the paradigm right. And it's clear to me that whoever is in the lead, because again, I think it'll change. Whoever's in the lead, it feels very obvious to me that all of the work that you do is going to be in one of those surfaces where, for example, when I'm writing a document, Codex has a browser in the app. It has an in-app browser. And when I'm writing a document, I just go into one of my Codex threads, which I have one thread for every project, and I just open the in-app browser, I go to the document, I usually do it in Proof, which is this online markdown editor I built. And then I just have Codex running and watching me in Proof and Codex can see what I'm doing. I can see what Codex is doing. It's all in one place, which is an extension of the same thing that made Claude Code work really well originally. And I basically feel like I have this parallel work buddy that not only can it respond and write in the document, but then it can go do research, it can use my computer to basically do anything that I can do on my computer and that's incredibly powerful. And I do this with everything. I've been at inbox zero for 10 days straight now, which if you know me, is crazy. I'm never like this, and that's because I literally just have Codex gather all my emails with Cora, which is our email agent. And then it renders a little page, and I think I showed you this at the Anthropic event. It renders a little page and I just monologue into it and just talk out each email. I'm like, "Okay, go research this. Oh, here's a question from our lawyers. Can you go collect all of the documents from the last four years and then put them into a report and send them?" And it just does it. And so all the stuff that I would procrastinate on, I don't really procrastinate on anymore. And so I feel like there's this, for a long time, I thought too that the optimal experience of AI was going to be take AI and put it in a browser.
00:23:11 · Dan Shipper, Lenny Rachitsky
And I think the reverse is actually starting to happen and be really, really valuable in a way that I did not expect, which is take the AI agent that you use all the time on your computer and put a browser in it so it can see everything you're doing. And that is just like a magical combination that I think is very uncommon now. You can't even do this in Claude Code because they don't let you browse external websites inside of Claude Code. So it's very uncommon now, but I think it will be super common in a year. [00:23:45] S0: This is more profound than it may even sound. What I'm hearing is instead of AI being baked into SaaS tools, what you're predicting here is the SaaS tools will run within Codex or Claude Code? [00:24:02] S1: That is one- [00:24:00] S0: ... code.