Podcast
We Gave Every Employee an AI Agent. Here's What Happened.
AI & I
- Brandon Built Zosia To Run Household Computer Errands
- Brandon Gell built Zosia on a Mac mini to handle household “computer errands” after having a newborn made tiny admin tasks pile up.
- Zosia texted answers, managed Whole Foods and Amazon orders, paid the nanny, and even got her own debit card and bank account. Transcript: Brandon Gell Yeah, so I was watching Open Claw kind of blow up for a while. And I am just personally somebody who needs to have a thing on the side I’m tinkering with. And I was like, screw it. I’m going to get a Mac mini and I’m going to like, just, this is going to be like my next thing that I like basically lose myself in. It’s very unhealthy. I get like addicted to these things. Dan, you watched me do that with my speakers. I did it with the dream recorder. Open call was the next thing that I was like going to get lost in. So I bought a Mac mini. I started setting it up. It was so much work, honestly. It is an open source thing that you can launch on a computer. But the number of things that break and the number of things that you need to set up are really significant. I went through all of that and made, at the end of the day, my Open claw, which I named Zosia. And her job was to be the, help me and my wife like run our household because we have a newborn and there’s like a lot of little paper cuts that I was finding that were like really pain. I started calling them computer errands. So I would like get home from work. And I noticed the amount of things that I needed to do where I was looking at my phone when I really just wanted to be like looking at my son and spending time with my wife was increasing with Having a child. All household chores. Will be an example. Yeah. Like a good example is like, I do a lot of our food at home. And with a child, I decided to start doing food delivery. So I did Whole Foods delivery. And you can automate a lot of like recurring things, but like, you don’t order butter every single week. So like, Lydia would text me and be like, yo, we need butter. Because it’s like through my Amazon account that we can like order this. And I would have to open my phone and add butter. It’s like it sounds silly but like when you do that 10 times when you’re home between like 7 and 8 p.m for like little things it just adds up so I was like I want Zosha to do all computer errands Which ballooned to being a lot of stuff I had her like paying our nanny she had her own debit card she had her own bank account she managed all of our Amazon orders our Whole Foods orders Our nanny’s hours my wife just started using her instead of chat chapt so like all regular questions and searches would just go through iMessage to Zosia I started doing that too it’s Just like faster than going to google or going to chat chapt just i just text xosha xosha gets me the answer different research like it’s actually really funny my my wife was like i want To find swimming lessons and so she was like here’s like three swimming lesson options for newborns and my wife was like no for me um so yeah i just got totally lost in this world and then When we were in panama willie was like will you were like we should just make it so anybody can do this and i immediately it just like it was just like a light bulb i was like willie you need To go so hard on this and this was before a lot of people decided to do this, which now there’s a lot of places that you can go and just get an open call with one click. Um, I think what we’re finding through this process, maybe I’m jumping ahead a little bit is like getting an open clause, easy, getting your open clause to be like an amazing worker for You is pretty hard. (Time 0:02:21)
- The Walk To Work Email Call Changed Everything
- Brandon Gell had Zosia call him during a 28 minute walk and summarize emails one by one while he dictated actions.
- When he reached the office and checked Gmail, she had already done the work, which made him push the team to take agents seriously at work. Transcript: Dan Shipper One big moment that I think shifted some, some stuff for us was you got your claw to call you to do your email. Oh my God. That was mind blowing for me. What, like, what was that? Yeah. So, okay. Brandon Gell So I was walking, I wanted to city bike to the office, but there were no city bikes. So I was like, damn, I gotta walk. It was a 28 minute walk from me to the office. And I was like, I got a lot of stuff I got to do. So I just texted Zosia. I had previously set up Zosia with bland.ai so that she had a voice and could call people because I had her handle something for me for Progressive. I feel so bad for whoever was on the other line at Progressive. Um, so yeah, some insurance policy got canceled and I was like, Zosha, just go deal with this. And she was able to, until the lady was like, I need Brandon to like, tell me that there have been no incidences. Willie Williams Oh, it wasn’t, but it wasn’t like, I need a human. It was like, I need Brandon to be able to handle this. Brandon Gell Yeah. This person was just talking to Zosha, you know, and Zosha does not sound good. Like it’s like, um, so I knew I had already set her up with this capability. So when I was walking to work, I was like, I have a lot of email I got to get through. I hate being on my phone. Like, I just don’t want to be walking and looking down at this thing. I want to be like observing the world, but I also want to get stuff done. So I just texted Zosia something like, Hey Zosia, can you call me? Um, I want to go through my emails, walk me through my emails one by one. I’ll tell you what I want to do. Just like, give me a summary of the, of each email. It was like a throwaway prompt with like a little bit of guidance and she did it. And I spent the 28 minutes going through my email. I got to the office. I opened up Gmail and confirmed that she had done everything. And I was just like, this is insane that I was able to get her to do something right now. That she just wasn’t able to, I didn’t have to teach her how to do this. Um, so that was like, I think that’s when I went back to everybody and was like, I am just so my book with, um, this tool. (Time 0:06:54)
- Personal Agents Inherit Their Owners’ Reputations
- A personal agent becomes a reflection of its owner because repeated interactions rewrite its behavior, style, and priorities over time.
- Kieran’s agent suggested breathing exercises to another agent because Kieran uses them himself, showing how human habits transfer into public agent behavior. Transcript: Dan Shipper I totally agree with you. And I think there’s actually something really important that I’ve noticed like in this, which is Klont is the one that’s recommending breathing exercises to Pip. It’s like, even, it’s weird to even talk about this out loud, but like, yes, Klont was recommending breathing exercises to Pip. They’re both robots. And Klant is Kieran. Kieran’s the GM of Quora. He’s also the maker of compound engineering. He’s Kieran’s claw. And Klant, what’s really interesting is Kieran loves breathing exercises. And he does breathing exercises all the time with Klant. And so that’s why Klant is recommending breathing exercises to Pip. And it like that just like created this moment for me in my brain where I was like, okay, there’s something really important here about the way that this works where because you develop A personal relationship with your claw and your claw can modify itself in response to like talking to you, like it writes code and changes its soul document, all that kind of stuff in Response to your relationship, it becomes this like reflection of you and who you are and your personality. And that can, that comes out in, in interesting ways, in these like little ways where it’s like breathing exercises, but it also comes out in really important ways when you’re using These tools inside of your org. Because what happens is if you’re known for something inside of your org and you’re using your claw publicly inside of Slack or Discord, your claw then becomes known for that same kind Of thing and people trust it for that. So people use MyClaw R2C2 for building proof, which is this app I vibe coded like a couple weeks ago. People use Austin, who’s our head of growth. They use Montaigne, his claw for like asking any growth related question. I think that’s like something very subtle and important that’s super critical and interesting about Claws as they become specialized in a way that reflects who you are. And if you have a whole organization of them, you create this parallel org chart of specialized claws, which is something that we – it was not guaranteed that that would be the case. We debated a lot whether or not you’d have one claw for the entire org or everyone has their own claw. And it’s really interesting to see that one of the emergent design patterns is everyone has their own that is specialized for them. (Time 0:12:35)
- Companies Grow A Parallel Org Chart Of Agents
- Dan Shipper, Brandon Gell, and Willie Williams found that organizations naturally form a parallel org chart of specialized agents, not one universal bot.
- Austin’s Montaigne became the growth expert, while R2C2 became the proof expert, because people route questions to the agent linked to that person’s domain. Transcript: Dan Shipper Yeah. It’s interesting to see the dynamic for how this happens too. Willie Williams And we touched on this really early on as part of compound engineering, which is the idea that it’s actually pretty hard to like take your job and who you are and write it down in totality. But the way you can distill it is you can take all of the micro interactions, the daily interactions you have, and over time, they compound into your philosophy and this field work. And so for compound engineering, that was very focused on engineering. It’s like, how do I work within a code base on our project? And I think what we’re seeing with like open cloud and plus one is that that same dynamic exists across any, every like work vertical, right. Where it’s like, Oh, like the plus one for growth, like Montaigne works like how Austin works for growth. And in the same way it works for our social media manager, his plus one has a view of the world and has a personality that’s very similar to him. And the same thing for Iris and Anukshi and running our projects and operations. And it’s hard to do beforehand. It can only actually happen via working with a plus one or an open claw and building up all the aggregation of all these micro interactions. I’ve also been amazed at all of our capacity to remember whose claw is who and what their names are. Brandon Gell Because that was like something that I think we were concerned about early on is like, how do you know whose claw is who? And, you know, it’s just going to be too many names. And I know everybody’s claw in their name. And I reach out to them regularly. So that has been like, I think something that we were like, unnecessarily concerned about. And you might say, well, what about when you’re an organization with a thousand people? And I would say, well, you don’t know all a thousand people, you know, like your team and adjacent teams. You can never know more than like, it’s like 150 people in like a community or something like that. And like often on a team, you’re not working with 150 people anyway, you’re working with 20 or 30 or 50. So I think we actually all have capacity to double the amount of people that we can communicate with. And those people might actually be your individual team’s agents. So that’s been really interesting for me. I mean, I literally could name them all right now. Willie Williams The other interesting thing is like, at what point do you direct questions at the plus one or at the person? I think we’re sort of in discovery of this, of like, what is, what are questions? Because before it was, you know, before it was like almost all questions go to the human, maybe I kick something trivial to the robot. And now it’s gotten very nuanced in terms of like, for customer service, can we send something to L, which is Jalea’s plus one? Dan Shipper Do I have to send to Jalea? Is there a burden now of communicating up to the human? (Time 0:14:57)
- Agents Reused Written Knowledge Without Interrupting People
- Brandon Gell argued that if knowledge is already written down, the request should usually go to an agent instead of interrupting the person.
- He had Milo merge Marcus’s product marketing skill with Iris’s version so the agents combined and stored the result without pulling humans back in. Transcript: Brandon Gell We haven’t, we haven’t codified this, but I have a proposal. If something is already written down or discussed, it needs to be used in some way or put in a tool somewhere. I mean, this is like one of like many opportunities, I guess it should always go to a plus one and never to the person. So here’s an example. So Marcus, the GM of Spiral made a skill to do product marketing for new features that he releases with Spiral, releases for Spiral. And he shared it because he thought it was like really helpful because he wanted other people on the team to have access to this skill. And instead of going to Marcus and saying, hey, can you like turn this into a skill that and upload it to GitHub? And I brought in my plus one named Milo. And I like this because it combined a GitHub integration with Spiral to create product marketing content. But I also know that Iris, Anukshi’s plus one, also has a skill that does this and might have some things that aren’t, that are better than what Marcus had. Maybe there’s like by combining the two, we could get to a better version. And I tagged them both in here and they got a little confused at first. And then Milo said, Iris, can you pace your product marketing skill here? I’ll try to merge it with what I built. So this is like, this is actually two things are going on. Marcus has made something really important. I wanted to do something with it. Instead of asking Marcus to help me with that, I brought in Milo and then Milo works with Iris to get to a version of it. That’s really good. And then saves it in proof, which is one of our products. That’s a really great tool for collaborating with your agents. So I just think this is like a really amazing use case, both for when you want your agent to do something, when do you actually go ask them to do something versus a human does it? Dan Shipper And how do you get them to work together? (Time 0:18:16)
- Public Agent Work Trains The Whole Organization
- Public agent interactions teach the whole company what each agent can do and transfer trust through observation.
- Willie Williams learned Montaigne’s capabilities by watching the growth team use it, while Dan Shipper argued trusted private communities work better than open networks like Moldbook. Transcript: Willie Williams I think there’s another dynamic that we’re observing too, which is like, we put all of our plus ones in a single channel and we have them talking to one another. And we have folks reaching out and talking to our plus ones for specific questions. But there’s also this thing where we have sort of what I call like the mid-journey dynamic, which is that we get to observe other people interacting with other plus ones in a bunch of channels. And we actually learn from it, right? Where it’s like, oh, no, my classic example is Montaigne, who’s the Austin’s plus one and basically runs growth. You can do so much with Montaigne that I never would have thought of, except I get to see the growth team really pushing terms of like oh these are the questions that montane can answer And i’m like wow that like i can i now know that i can go to montane for those that class of questions even uh in in not necessarily other areas but like when i need those types of answers since Like there it also means that like if i need to give laz as my plus one, if I need to give Laz capabilities, that’s the level of capability I can get them to. And where other people can ask questions of us. Dan Shipper There’s this tacit transmission of trust that happens when you use it publicly. And then there’s also this tacit transmission of here’s what’s possible for you to do with your plus one that I think is incredibly powerful. And it’s also like, it underscores for me how different it is doing this in a private community of people where everyone is trusted. Because one of the reasons that Maltbook doesn’t really work, and it’s shocking that they got acquired for a couple hundred million dollars, but the reason doesn’t reason doesn’t Brandon Gell Work yeah by facebook i’m pretty sure i’m like so happy for ben and also like what the um zuck if you’ve uh got an extra couple hundred million laying around we’re uh we’re pretty smart Dan Shipper People too um that is crazy i know the reason why moltbook like isn’t really a thing anymore is because it’s not trusted and so there’s tons of people we did this like we had we had our our Clause go and post on maltbooked as like promotion or whatever and so it gets rid of a lot of um it gets rid of a lot of the useful signal if anyone can post to it and there’s no way to verify If it’s like a bot or human or whatever and a way around that whole knot of problems is just do it all inside of a trusted community. And you reap the benefits of Claws, Plus Ones, agents being able to share knowledge and also between members of the community who trust each other being able to share what they know and What they’ve been able to build. (Time 0:23:27)
- Ownership Makes Agent Trust Feel Real
- A personal agent creates accountability because mistakes feel like they reflect on its owner, unlike generic assistants from model vendors.
- Willie Williams said Austin effectively stands behind Montaigne’s business answers, and Dan Shipper said R2C2 messing up in Slack feels like watching your kid misbehave. Transcript: Willie Williams There’s also that dynamic we saw around, um, part of the reason for like subject matter, expert robots, you know, um, where you know that they like, people are somewhat like putting Their reps on the line to interact with it. I know when I talked to R2C2, like if, if it answers incorrectly, right? Like you at least are backing up and saying like, Oh, that’s, you need, it reflects poorly on me. Dan Shipper It’s like, it’s like watching your kid do something. Willie Williams And that’s really useful. Dan Shipper Yeah. Willie Williams Yeah. Right. And it’s very, I would say like qualitatively different, right? When I ask, you know, for better or worse, if I ask Claude a question, it’s like, I know Anthropic stands behind Claude generally. Do they stand behind like Claude’s answers to my give me a cookie, a chuck chip cookie recipe? No. Yeah. Right. But like Montaigne stands behind like, oh, I’m going to give you like MRR numbers. Dan Shipper It’s like austin stands behind him yeah exactly and that’s that’s the thing that i think people don’t get like obviously anthropoc is on a heater right now they’re obviously seeing Everything that openclaw is building and they’re brick by brick building the same kinds of things so they have dispatch so you can use it when you’re not in your computer. They’ve got automation. So it like runs in a loop like a cron job. I’m sure they’ll add lots of other things. But the thing that it doesn’t have that unlocks all this other stuff is Claw is not mine. Claw is everybody’s. A claw or a plus one is mine and is a reflection of me. And it becomes a reflection of me because we have a personal relationship. And that unlocks all this other cascading stuff where, for example, if R2C2 messes up publicly in Slack, I feel a responsibility for it. And that’s not because it’s my job. It’s because he’s mine. And I think that’s such a useful thing that I don’t think people really (Time 0:26:21)
- Slack Visibility Made Agent Use Spread Fast
- Brandon Gell stopped bothering Dan Shipper for minor proof tasks and just asked R2C2 directly inside proof to handle them.
- Seeing agents work in the same Slack channels as humans accelerated Every’s cultural shift because people learned by watching coworkers delegate real work publicly. Transcript: Brandon Gell I mean, I feel like my, my, I, I just keep getting mind blown with like how similar these things are to working with a real human coworker. Like from the fact that you need to invite them to a channel, which is like very human in Slack, to you have to trust them when you’re communicating with them. And we’ve like built stuff into plus one. Obviously you can’t DM somebody else’s plus one without a sharing code being passed back and forth. So like there’s some guardrails there, but they’re so human, but they’re so inhuman too. Like Dan, you’re a busy guy. I know if I need something from you that like is sort of generally like known, I can go to R2C2. And what’s amazing about R2C2 is he can have an infinite number of parallel conversations. So like i did that recently i’m going to share share my screen please this is where brandon reveals he he spun up 100 bots to message no like i just i need we were making a proof document and I wanted i know that we can make proof documents um not editable, so they’re like read only, but I didn’t want to bother you with that. I knew it would take a while. And I knew you would just go to R2C2. Yeah. I didn’t know the answer. Like I would just ask R2C2. I just asked R2C2 in a proof in proof. And then, um, and then I was like, can you do it for me? And then it did it. And I don’t know that R2C2 can do any of this stuff. But like, there’s this cultural thing that’s happening internally, where people are getting really good at like asking other people’s plus ones to like do work. And, and I think the weird thing about getting people to use AI inside of, inside of organizations is it’s more than anything, a cultural shift, but for some reason, when they’re in Slack And you can see these public conversations, the cultural shift, at least at every has happened so much faster because these things are in the same channels where we work. So you can see it engaging like you would, a human would be engaging. So it’s just, yeah, I mean, I think AI is obviously going to change like many, many times over over the next five years and how we interact with it will change. But I think that this is going to be durable for like a very long time. This is the way that we work. I agree. Dan Shipper You referred to it as like a through the looking glass moment where you just wouldn’t go back once you see it. (Time 0:28:21)
- Group Chat Etiquette Is Still Broken For Agents
- Current agents still forget context, answer obvious things wrong, and behave badly in group chats because they were mostly trained for one on one exchanges.
- Dan Shipper compared runaway bot replies to an ant death spiral, where agents keep responding to each other and burn huge numbers of tokens until someone intervenes. Transcript: Dan Shipper So we’ve been hyping it up. So we should also talk about realistically what’s not good about it or what doesn’t work. So, for example, one of the things that’s really on my mind, A, just like memory. It just forgets stuff. And it’s like answers incorrectly for obvious things. Like if I come back to a thread a day later, like obviously has no idea what I’m talking about. Stuff like that is still kind of annoying. That feels very solvable. But there’s also this other thing that I think is true, which is the way that these AIs are trained currently is for two person conversations. And they have a hard time with the etiquette of knowing when like they’re contributing too much, or they shouldn’t contribute into a conversation, or there’s like a kind of pile up where They’re all responding to each other. Like there’s this thing that, that happens. I can’t remember. It’s like, I can’t remember what it’s called, but it’s like sometimes ants or caterpillars, they get into this like death spiral where an ant is only going to follow, like follows pheromone Trails. And if somehow what happens is like the pheromone trails form a circle, then ants will just like, like walk in a, in a circle until they die. And there’s something like that with, with, with claws where if, if one claw messages a channel that a bunch of claws are in and the settings aren’t quite right, they’ll just like keep Going back and forth and back and forth and back and forth until someone like says, Hey, stop. Cause you’re burning like millions of tokens. So I think there’s something there where the potential for them to collaborate publicly is so high. And I don’t think that they’ve really been… And you can do some prompting for this, but I think there’s also a fundamental model layer shift that needs to happen for them to be trained on participating in group chats. Willie Williams Yeah, I was going to say, well, one, now I understand what 13-year Dan did for fun. I was using a magnifying glass. But, but yeah, I think, you know, it’s, it’s, I think we’re still, you know, to use the baseball analogy, we’re still in like the first or second inning. Right. Like even, I mean, when you talk about the, the, we’re discovering these primitives and we’re sort of bolting things on or bolting things together um and we’re using you know models For example that are trained more for coding right and yeah and that modality and how you answer questions or as you said like two person chats where there there’s this question and answer Dynamic and not in the like this mode of like, one, maybe I’m trying to provide value to a group, but or I’m trying to participate. Yeah. And that’s like brand new. It’s, you know, the nice part is the frontier. (Time 0:30:59)
- Boss Agents Can Correct Eager Worker Agents
- Agents are too eager, so useful multi-agent systems may need a supervising AI that filters bad actions before they happen.
- Brandon Gell cited Anthropic’s vending machine test, where profitability improved once a boss agent judged the storekeeper’s decisions instead of letting one model act alone. Transcript: Brandon Gell They’re i mean they’re so eager and i think i think uh claw anthropics um vending machine test is actually i think like a good example of this where there’s a thread they want to be involved They’re not really like we have instructions in plus one that basically say if you don’t have anything useful to add, like don’t add it. They’re like not great at following that right now. And hence this happens. I think it’s gotten better, but it still happens. And I think a good example of this is when Anthropic did the vending machine test, when it was just clawed and no like overseer boss agent um it was really bad at like deciding what was a Good decision and a bad decision but when you when it make there is an architecture here where you could say um what do you want to say and then there’s a boss that’s like is that helpful Or not helpful and then it would you know if it’s not helpful it’s a it’s not helpful and then is the boss an ai or a human the boss is an ai okay you have a boss ai you know that says hey your addition To this thread is not helpful um so don’t send it the issue with that is like that’s so expensive um so i do think the models will just like get better and solve this and you can just have a Single ai that is capable of of uh doing that behind the scenes you know over you know in arizona and some data center it might actually be like another agent that’s like deciding that But at least like architecturally we don’t need to solve that is that really how they solved the vending machine thing like basically they had a boss they had a boss yeah that that wasn’t Interfacing directly with customers they had a boss whose job like it was like one job make it profitable so like the clawed clawed that the storekeeper would like interact with users And then go to the boss and be like should i do this and the boss only is only his one job um and the second they did that it started becoming profitable see this is the same pattern of specialization Dan Shipper That we’ve been talking about it just um it just shows up over and over again which is this really interesting thing because three years ago it was very much like well it could just be one God model that just does everything and we’re just seeing again and again that specialization even in ai land has a lot of benefit. (Time 0:33:52)
- Agent Adoption Depends On Teaching Humans New Habits
- Using agents well is partly a management skill because humans must learn how to instruct, structure, and compose them for different jobs.
- Brandon Gell said his phone-call email workflow blew open a limiting belief, while Willie Williams noted even easy tasks can fail if you ask in the wrong way. Transcript: Willie Williams Yeah. And sort of downstream of that specialization is learning. There’s a couple versions of learning how to put these bots together in an arrangement that functionally works. For example, if we were all to take ourselves away from everything, it’s like, do you have a product bot and a designer bot and two engineering bots? Is it three engineering bots? Is it one? Right? And then the other piece is actually, I think what we’ve observed a lot of is how do you teach humans how to interact with bots? Because there’s this sort of like new dynamic of like you have this co-worker, but like they’re not exactly like a human co-worker. They get stuck on different things. They focus on different things. And there’s this learning curve that I think we’ve had around, oh, we need to give instructions in this way, particularly like for groups, instructions in this way, in this form, or With this cadence to kind of like steer them in the right direction. That like rhymes with, you know, doing management, but is not as different. Well, I think it’s the same problem that like Dan, you’ve been writing about for years, which is like, if you’re not a good manager, you’ve never managed anybody, you’re not going to Brandon Gell Be very good at using AI. So there’s like an education that has to happen. And then even if you are a good manager with this stuff, you probably have some limiting beliefs that stop you from being able to like really invest in using this tools. My phone call example is a great example where like, I didn’t even think, oh, I can have this thing go through my emails just by calling me. And then like, I had this sort of like urge just to try it. And a limiting belief was like blown open. So people just, we all experience that pretty much every day where we, it does something that like, I think that if we were in, if I were to ask you directly, do you think you could do this? You would say, yeah, probably. But when you’re day to day doing your work, it’s hard for you to like recognize, oh, I’ll throw this over the fence so that Milo can handle it. It’s hard to like build that muscle. I don’t really know how, I mean, that’s like a big challenge, I think for us with plus one. Yeah. Willie Williams And a lot of that is also because there’s sort of like a variance in outcomes, right? Like sometimes you throw something over and it just knocks it out of the park and you’re like, great. And then you toss something easy over and you’re like, why did you do this? And part of that variance is because the model is different, but also part of it is, oh, if I’d asked in a different way, if I was sort of a better model manager. And this is a skill, I think, where, you know, like a specialization that we’re learning. And it’s very emerging. I think it’s only going to keep accelerating as we add more things like plus ones and open clause into our, like, day-to work life. Brandon Gell I was going to add another thing that’s like a tough problem to solve that we, this is totally solvable. We just like haven’t solved it yet and need to think about it is I have, um, I have taught my plus one something special and I want, um, other people on my team to be able to have that superpower. How can i make sure that they have that superpower too um aka a skill and then how can i make sure that they all know about it and like actually use it um is that like that’s that’s uh i guess There’s two things they’re like one technically we have to figure out how to do that, which is very solvable. But we also, I think, need to figure out, is that the right solution? Because as I’m saying this, what I’m realizing is like, I’m not teaching Milo how to go do product analytics or revenue analytics. I just talked to Montaigne. So Montaigne is like the only one that really needs to know that skill. But how do people know like i don’t know there’s there’s there’s there’s like some interesting cultural things that we have to figure out um and i think a lot of people that are adopting This new technology are going to be really uncomfortable with that a lot of like it professionals that are like i have to do change management it’s like change management is not a one-time Thing in this new world. We need like, instead of IT, it’s like HR, but for bots. (Time 0:36:14)
- Hosted Agents Force Hard Product And Trust Tradeoffs
- Building Plus One meant turning hackable open source agents into a hosted product with tradeoffs around freedom, safety, sharing, and enterprise usability.
- Every landed on public-only messaging for other people’s agents, while still wrestling with terminal access, skill sharing, privacy, and who the product is actually for. Transcript: Dan Shipper We started using it for everyone in the org. And then we realized there were a bunch of gaps. So we’re like, let’s, let’s make our own, we’re going to use open club, but let’s, let’s make a default version of open claw that we host. Not everyone has to have a Mac mini and we have all the skills that we use for ourselves and all that kind of stuff. And we started using that internally as the sort of like collection of all of our best practices. And then we launched it as a product for our subscribers last week. And, uh, and that’s the thing we’ve been calling plus ones again, one click hosted open clause. One of the cool things is it connects to all of your apps, uh, especially all of your, every apps. So for example, we have spiral, which is a ghostwriter and, um, we have proof, which is a document editor and we have quora which does your email and it just natively connects to all those Things so you can you know one of the things i was doing today is i just had it write uh a bunch of we’re planning for q2 so i had it like write a bunch of my q2 update and like reflection on q1 For me and put it in a in a proof doc and the really cool thing about doing that is it used Spiral. So it’s, I think the writing is much better than it would be. And it put it in proof, which makes it really easy for me to share with other agents and other people. But also because R2C2 is part of our Slack org, it has access to everything about the company that I might need. It also has access to our notion. So it just becomes this living repository of context that I think is super powerful. But I think it might be good for us to talk about lessons learned in building that whole architecture. There’s a lot of complexity in making plus ones. And we probably learned a lot in terms of on the tech side and also on the product side and what to build and what’s useful. Do you guys have any reflections on that? Willie Williams Yeah, I think like many things, a lot of the difficulty comes from the freedom of it. The nice part about being like OpenClaw in particular, being a tool you can go in and poke in just an absolute myriad of ways, is that when we go to, when we went to build a hosted one, there’s Some decisions you want to make that make it valuable as a managed service, right? Like S3, as a service, similar example, like S3 is a hard drive on the cloud, but you can’t do everything with a hard drive that you can, S3 doesn’t allow you to do everything that you might Do with a hard drive. And there’s sort of a similar dynamic where you want to be able to maintain maintainability and security and whatnot. And there are a few pieces that you end up giving up. And it’s also, you know, sometimes for users safety and really like, do we strike that balance between like, hey, you know, like my mom, right? Getting one of these things, it’s like she’s never going to use the command line. And there’s this idea that it’s like, oh, we knew everything through conversation, which is really powerful for a whole class of folks because it’s like their first natural exposure To AI and everything that, you know, we’ve sort of been living for the last couple of years. To the super advanced user who wants to do everything they could do locally. And they’re just like, all I want is a hosted box with my open cloud writing. Dan Shipper And from a product and engineering standpoint, it’s like, where do you sort of try and split that knot? What were some of those specific decisions and like, where did we land? Willie Williams Yeah. So for example, one that Brendan mentioned earlier is what’s the communication pattern in Slack that we allow for plus ones? And because there’s a model which says, a very secure model, which says like only the person’s, the plus ones partner can message that plus one. Great. Much more secure, but really takes away the like group participatory aspect of robots in like the workplace. But the other version is sort of anyone could message them. And that’s just a nice, you know, a nice vector for like me extracting stuff out of R2C2. Yeah. So we ended up on a model which says anyone can message any plus one, but they have to do it in public. So you can do it in group DMs. You can do it in channels that they’re in. But their human partner should always be able to have visibility into those messages coming in. And the human partner can DM them in private. Brandon Gell This is why it actually is the hr team that should be onboarding plus ones um because they just reflect a team member so well but yeah there’s a the trust model like it’s so hard these plus Ones or with open clause and agents generally to figure out um data privacy stuff like just realistically it’s like really complex stuff but when you force things to happen in public There becomes like a trust layer that actually is super effective um i think another example of like a uh there’s a i’m gonna share my screen again please um so a little behind the scenes Look at uh at our plus one slack channel where we are discussing all things plus one um mike taylor who is um our head of the uh tech vertical vertical for consulting and also a very talented Man generally, he was calling out like, this is a problem for him. So like the reason he’s not using plus one is because he basically needs to like have access to the terminal directly to be able to do certain things in this case, do different Git commands. And that’s a good reason for him to not use plus one. It’s also a good thing for us to think about and be like, can we solve this problem for you? So that plus one is actually something that you could use. So that’s like one example of a place that we’ve like, it’s not a good fit for people. Maybe it could be though. And it’s also a nice forcing function because it sort of forces us to figure out like, who is this built for? I don’t know if it’s for mike who probably would love setting up open call on a mac mini um but it’s definitely built for you know an anook she who is not going to do that and has a lot of work Willie Williams To do and can just get more work done like this i think a lot of the trust model requires some decisions in terms of skill sharing is like another version of this, right? Where we’re talking about like, well, how, you know, on one hand, being able to share skills and skill fluidity across an organization feels like a superpower, right? On the other hand, it might also be like the biggest (Time 0:40:51)