Transcript: 'Why Every Traded Personal Agents for One Company Agent'

The transcript of The Every Podcast with Willie Williams is below. Watch on X or YouTube, or listen on Spotify or Apple Podcasts.

Was this newsletter forwarded to you? Sign up to get it in your inbox.

Timestamps

Introduction: 00:00:00
How OpenClaw spread through Every: 00:05:58
Why everyone’s personal agents eventually died: 00:11:13
The difference between building for builders and building for operators: 00:14:51
Why work will run on company agents while personal agents will stay personal: 00:16:20
How engineering management has changed in the age of AI: 00:19:56
Using an AI-powered feed to see everything happening across a product launch: 00:21:43
Why Willie’s token usage fell off a cliff: 00:23:53
The benefit of having separate AI coaches for each part of your life: 00:29:05
How Every is building its agent to surface information unprompted: 00:31:45

Transcript

(00:00:00)

Dan Shipper

The hot thing right now is Dots. They’re persistent agents from OpenAI. I was there at DevDay when they launched them, I interviewed Sam about them, and I’d been using Dots before they came out. Now we’re into week two of Dots, so how’s it going? If you read my Vibe Check, you know that I thought they were a little bit rough to begin with, but that the idea of a persistent, proactive agent was obviously the future.

Even before Dots launched, they’d become my main way of interacting with ChatGPT, and they’re still that way after launch. I’m really excited about the little things they’ve been doing to make the experience less buggy and more complete, and I think it’ll get even better over the next couple of weeks.

What’s been really interesting, though, is seeing the team at Every—we’ve got about 30 people—start to adopt them, and seeing some of the use cases that have started to emerge. So in case you’re looking for some use cases, I actually had my Dot collect some to talk to you about.

One thing Dots are really useful for is keeping up with school emails. Kieran Klaassen, who works at Every, says his scans all of the emails he gets from his kids’ school and tells him what he needs to know.

Another thing Dots are really good for is catching access requests, or little messages you might have missed in Slack, and making sure you get them. Arielle Schipper, our head of operations—her Dot caught a few people asking for different accounts across the company and asked if she wanted it to just approve access.

Dots are also really good for working when you’re not at your computer or your hands are full. Our head of video, Randy Kinsman, was running around yesterday trying to get our launch video out for the Every agent. And he was doing work for that launch just on the phone with his Dot, talking it through different video edits and the different things he needed to get done.

This is something that came up in my interview with Sam, and I feel it too: Dots are really good at filtering out all of the noise you don’t want to see and only showing you stuff that really matters. For example, I was doing a lot of model testing over the weekend, and normally model testing requires me to be in the Every Slack. Once I’m in Slack, it’s a rabbit hole—I’m just going down, doing tons and tons of work. What was amazing about using my Dot for model testing was that I was able to post the results of all the experiments I was running in Slack, and if someone responded, my Dot would tell me about the response. That helped me stay up to date on what the team was thinking without getting sucked into the rabbit hole of everything going on in Slack. It kept my weekend pretty clear and calm, and I think that’s a huge thing for Dots.

So what’s the overall vibe? If you don’t have a Dot, I highly recommend you get one, especially if you’re in the ChatGPT ecosystem. If you’re using Muse or Instinct or Grok bots, you should check it out. I think they’re pretty much at feature parity right now, but there’s something really good about having something that’s hooked into the entire rest of the ecosystem you use for your work and your life.

There are still rough edges, though. The whole permission structure of Dots is completely a mystery to me and is extremely annoying. For example, I often have to give my Dot multiple approvals to do something basic, and even when I give it approval, it says it can’t do it. They really need to figure that stuff out to make the experience awesome.

But I would expect that in a year, this is the default way you’re interacting with all your AIs. My other little hot take is that it’s probably not going to be a cute character. Mine is named Boo. He’s super cute, but my experience using OpenClaws and other named agents like this is that as soon as something better comes along, you throw away the cute little character you developed. So my guess is I’ll be using a persistent agent in a year, but Boo will probably be dead. RIP, Boo.

We’ve also been working on our own persistent agent. It’s not a personal agent—it’s a company agent. That came from this whole journey we’ve been on, starting with OpenClaws inside of Every, where everybody had a personal agent, and then realizing how valuable it was for the company to have a company agent. So we built one for ourselves, and then we just released it to all of our subscribers.

So I brought on Willie Williams, our head of platform, who was instrumental in bringing the agent to life, to talk about personal agents versus company agents, what it’s like to build an agent at the frontier right now, and what that means for engineering teams going forward. I hope you enjoy.

Dan Shipper

Willie, welcome to the show.

Willie Williams

Thank you for having me.

Dan Shipper

So for people who don’t know, you are the head of platform. You are one of the driving forces behind the Every agent, which, by the time this comes out, we’ll have launched—it’s launching tomorrow. And I want to talk to you about the whole agent phenomenon, because I think we’re in this world right now where personal agents are the hot thing.

There’s a little bit of Groundhog Day for me, and I assume for you, because we went through our personal agent phase eight or nine months ago with the original OpenClaw phase. I feel like we learned a lot from that. And now there’s this whole new era of people who are experiencing the same thing, and maybe they’ll learn the same lessons.

For us, we went from OpenClaw, where everyone in the company had a personal agent, to having a company agent. And we liked the company agent so much that we actually just made a company agent for other people to use. I want to talk about that as honestly as we can, because it’s possible we’re wrong. But I do think we’ve learned some things that are important. Maybe let’s start with OpenClaw, because we actually disagreed, and I think you ended up being right. So tell us about the OpenClaw situation and your take, and then we can talk about why I was wrong.

Willie Williams

Well, I think the funniest thing was, at the time, it was called Clawdbot. That was the earlier version, before it got clobbered. And even before that—

Dan Shipper

Moltbot.

Willie Williams

Moltbot. We forget about those two days. Big shout-out to Ash, actually, who I think first messaged me about it. Ash is one of our contributing writers. And he was like, “Dude, this thing is awesome.” And I remember because I knew Peter’s name from the top token burners leaderboard. He was up there, and I was like, “Oh man, these folks are burning a lot of tokens.” And then Ash was like, “Oh yeah, he’s doing this thing called Clawdbot.”

I was like, “Oh, this is kind of a weird thing,” and I kind of followed the project. And then Ash came back and was like, “Oh yeah, dude, this is really cool.” So we started trying it. And then I think Brandon was the first one who was like, “Okay, I’ve got one. I’m starting to wire it up right now.”

And then it really kicked off in January, because we went to Think Week. For folks who don’t know, that’s our semiannual gathering at Every, where we think about the future. And that was the place where we were finally all in person, like, “Oh, I can show you what’s going on.” I think this is a recurring theme: Showing people how to use AI is a very transformative thing, but telling them sort of lands flat. And so we had a bunch of people who were OpenClaw-pilled, who were like, “Oh yeah, this is cool. This is what I do. I order my groceries. I’ve got it in my group chat with my family.”

Dan Shipper

It really spread. We were all in Panama together, and I remember coming out of that—I think there was a model drop the day we left, and the claws just started working way better. And then I made this internal place for them to talk, and they started talking to each other in our Slack. And it was like, whoa, this is crazy.

Willie Williams

Yeah, it was a funny thing, where we were like, okay, the first blocker is that it’s just a little bit too technical. So we went through a couple of iterations of how to do it for the less technical members of the team. We had a template, so we were spinning up infrastructure for that. And then we made a little service where we could walk people through the setup and then they’d get their claw. But we were also looking at it like, “Oh man, security is hard for all of these things.” But I think we powered through a lot of that.

And I have this vivid memory of just so many bots in our Slack. Absolute chaos. Because you also had Brandon just having so many bots running around doing various things. And I think the big takeaway was, one, it was so valuable for people to see other people using AI. Up until that point, it was like, “Oh, I built this gigantic Claude or Codex command center on my local machine, and I’ve got it doing everything.” But the only thing people would see was the output of it. And once we had OpenClaws, it was like, “Oh, I didn’t know I could use it that way. That’s actually pretty cool.”

Dan Shipper

“And I didn’t know that you prompt it in this way.” There are all these little things people do when they’re using their AI that you never got to see. And it was like, oh, I’m learning every time I look in the Slack.

(00:10:00)

Willie Williams

Yeah. And the other thing that I think was really big was that we liked our personal bots. And I think this is another thing we’re seeing again now: You form an attachment. I was attached to mine and what it did, and how customized it was. So for folks who are tinkerers, builders who like playing with these things, these setups are great. For folks who are a little more on the operator side, it was like, “Oh, there are a bunch of security worries. It’s super complex. How do I just get it set up?” And if you don’t invest in it, it kind of fades. It doesn’t keep up with the rest of what you’re doing.

And that, for us, was the striking point. Instead of everyone investing in their own agents and making them better day to day over time, why don’t we take all that energy and put it into one agent that the whole company benefits from?

Dan Shipper

I think also what happened is everyone had OpenClaws, and for sure one of the coolest things was that your claw became a little bit of an extension of you. Brandon was a big driving force here, and the second person who really took it to a new level was Austin, who at the time was our head of growth. He had a growth claw that you could ask all your growth questions to, and that was really interesting.

But at least my memory of what happened is that we were like, “Oh, we’re going to build a platform for anyone to make their own claws, even if you’re not technical,” called Plus One. And I don’t think at that point it was really a company agent. It was just a one-click claw maker. And we tried that.

At the time, we were having a lot of internal debates about whether it should be one agent for everyone or an agent per person. You were very much one agent for everyone. I was very much agent per person. And I think you had a lot of hesitancy about building our own platform infrastructure to support this claw-as-a-service thing. And we were like, “Fuck it, let’s do it anyway.” And you were like—

Willie Williams

Okay.

Dan Shipper

Which I appreciate, but in retrospect, it was not the best idea. We launched that in beta, and we found that there was a lot of demand for it, especially at that time. But OpenClaw itself was very hard to build a platform on. It was moving so fast, and it was completely vibe coded. It’s just hard to build the entire technical stack when the thing you’re serving is an open-source product that’s not really designed to be a platform. So that was messy.

And through that process, I think all of our claws basically died, because we just stopped. I remember my claw’s name was R2-D2, and I had this really attached relationship to R2-D2. And then at some point, someone knocked out the Mac mini’s power, and it disappeared. And I didn’t restart it, because at that point I was using Codex and I was like, this is fine. I don’t need that.

That was a really interesting thing for me to realize: Okay, there are all these people in the org who don’t really want to maintain their own claws. And the people who do want to do it eventually get tired of it if there’s a better option. I think we were all starting to use Claude or Codex. And then there were a couple of company agents we tried that really changed how we thought about this—Victor being one, which was just really, really well designed. And I think at that point we started to realize that having a single company agent makes a huge difference.

Willie Williams

Yeah. Even just naming—remembering which agent had which expertise was annoying and hard. And I remember Austin’s agent, Montaigne, sort of became the gravitational center, because he’d built it up to be so good at the growth job. So it was like, “Well, I’m already using it myself for minor growth jobs. I could invest in my own agent, or I could just add on to that and get the gains from it.”

So I think there were a lot of pressures there. But of course, counteracting this, we really liked our personal agents, and we were trying to keep them. The other thing I think is underappreciated is that when you’re building AI products now, the surface area for products and agents is really large. Before, it was like, “Oh, I’ve got a SaaS app. I can map out all the user flows, go build it, and I sort of know how users will interact with my product.” In the AI world—and with Plus One, this is what we noticed—it was like, “Oh, we didn’t even realize you wanted to go do this thing. How do we support it?”

I talk a lot about building and managing complexity for builders versus operators, and it’s a hard job. Builders want to be exposed to the complexity, because everything they want to do is going in there, tinkering, and turning knobs. But if you’re trying to run a business, you don’t really want to do that. Mostly you just want something that works for you. And I think over time we realized that we wanted something that just worked. We didn’t want to deal with all the little knobs and bells and whistles. For the jobs we were using an agent for—a lot of product and project management, operations, and whatnot—it was better to have something solid.

Dan Shipper

Yeah. And to have it be shared, so when you use it, you’re pretty confident that it’s the right thing to be using for your company, and someone has thought about the skill or the feature or whatever. So now we’re in this place where we’re launching the Every agent, which we use many, many times a day—the company would probably kind of fall over if we didn’t have it—and we have a bunch of beta customers.

And we’re also in this moment where, for example, I’m using my Dot from OpenAI all the time, and there’s Muse and Instinct on the personal agent side. I’m curious about your bet here. One of the things I feel is that I will be using something like my Dot in Slack a year from now, specifically because it’s basically just a portal to my Codex. But mine is named Boo, and it probably won’t be named Boo. Boo is going to be dead. Its life span is maybe six months.

Because the personality thing is cool, but particularly in a work context, I mostly just care whether it does the thing. I have this personal connection, but the minute there’s something that performs better, even if it doesn’t have that, I’ll just switch. I think there’s much more room for that on the truly personal side. So my bet with Dots is that you still actually need a company agent, like the Every agent, and then that needs to play well with everyone having their own Dot or Dot-like thing that’s basically a portal to their command center. What do you think?

Willie Williams

Yeah, I think we can see the line between professional and personal shaping up very solidly. On the professional side, I do think it’s one agent per company or team structure, because there are just so many advantages to it learning your workflows and what you do. Everyone contributes just by working with it and making it better, so it shapes itself to your company. If you bring a personal agent in, it doesn’t have that context. It’s going to be hard to teach. It needs to have been there a year to really get up to speed. So I think we’re going to have that for the workplace. And I really don’t see us or the industry repeating the chaos of every worker getting their own AI coworker, where we all double our company size.

But on the personal side, I actually think the opposing force—personalization—is huge. The thing you love is that you have an agent that is just yours. Before, the downsides were: It was open source. How do I do security? How do I manage it? How do I run it? There were all these questions that prevented it from gaining larger adoption. And now that bigger players are coming in with more money, we’re starting to get closer to a better consumer personal agent experience, and it’s going to be incredibly attractive.

So I think we’re going to see a split. I go to work and work with my company agent, because we’re all working together toward a collective goal. And at home, it might just be me, or me and my family, using a personal agent to help us do very similar things, but it’s personalized to us.

(00:20:00)

Dan Shipper

One thing we’ve been talking about a lot, and that I think you have a very good perspective on, is this: You’ve had a long career in tech, doing engineering management at great startups and bigger companies. You’ve founded your own companies. And you led the engineering for this effort with a team of, I don’t know, five to seven-ish people. How has engineering management changed, especially over the last couple of months? And what have you learned in the process of bringing this to market that you didn’t necessarily know?

Willie Williams

Ironically, engineering management feels the most similar of any of the jobs I’ve done before, especially now that I’m getting to come back to it and see what it’s like in the age of AI. I think there are lots of advantages. For an engineering manager, a lot of your leverage comes from how much of the field you can see. In previous years, you would spend a lot of time trying to gather that information.

Dan Shipper

What would be an example of something on the field that you need to see?

Willie Williams

Oh, there are a lot of Slack channels—keeping up with Slack channels and emails and customer reports. How are you integrating all of this into a single place? I run a Tend feed that does this—it consolidates it all into a single feed. During the day, I can just go through it, and it gives me much better vision into everything going on related to the project. Is marketing happening over here? Are we seeing customer reports over there? Normally, that would have been meetings.

Dan Shipper

That makes sense. So for people listening who don’t know, Tend is an open-source app I made that lets you create custom feeds for yourself, where you hook up data sources—meetings, Notion notes, anything like that. And then your AI automatically turns those data sources into a feed that either has decisions you need to make or is like, “FYI, this thing happened in a meeting. Here’s what you want to know.” It learns over time what you want. So you have one for your engineering stuff, and I have one for the full company. What are a couple of things you’ve learned, maybe even in the last week, from reading it that you might not have known otherwise?

Willie Williams

A lot on the marketing side of launching a product. Normally, when you’re building a product, you’re in it with your teammates. You’re looking at screens and designs and bug reports and whatnot, and you have a feel for everything going on. But when you’re launching a product, there are a lot of conversations and meetings that are public, but they’re just out in places where you don’t necessarily have the time or the attention to check. So how do you grab those things and make sure that, one, they’re going the way you want them to, and two, you know who to give guidance to?

Sometimes the rumor mill of a company comes back, like, “Oh, did you hear we’re doing this sort of thing?” And it’s like, “Well, no, we shouldn’t be doing that. But who do I talk to?” That’s a huge question. And as a manager, that’s your leverage: I need to be able to go talk to people and set guidance and vision for all of this. So having this field-marshal-level view of a lot of things is just so helpful, even for a small team.

Dan Shipper

Yeah. I think you’ve been on the receiving end a few times of me being like, “Hey, I read in Tend that this happened in a meeting. What’s going on here?” Just to put my thumb on it. As a manager, it’s the coolest thing, and it can be weird for employees, but I’m very curious to see how that changes. Now would be the time when, as a manager—as an engineering manager—you could actually be ripping tokens while you’re managing a fairly large team. What’s your take on that?

Willie Williams

I think a lot of it comes down to people. The job of a manager is still to be a force multiplier on folks. So where do I spend my extra epsilon of energy? It’s really probably not on tokens for coding. This is the part that’s surprisingly very much the same: It’s still very much a people job, and you want it to be a people job. The difference is that your leverage has increased because their leverage has increased.

So the most valuable things for me recently have just been: What are the bottlenecks, and how do we go and remove them? It’s funny—as I’ve transitioned from doing more technical work back into engineering management, you can see my token usage just fall off a cliff. I’m using AI to gather all this information, but when it comes to action, there’s no billion-token move I can make. It’s mostly that we still need to sit down and have a conversation.

There’s a whole bunch of routing functions that are now helped by AI. We can do summaries—how do we summarize and pass along information? A lot of the jobs managers used to do, and a lot of the things they used to be responsible for in terms of ops and the health of the team and the company, are much less manual now.

Dan Shipper

Have you automated that?

Willie Williams

Mostly for the agent team, yeah. We’re good. It was just amazing to me, having built a lot of dashboards and alerting systems before. It’s really definable and really testable, and that’s the kind of thing agents are great at.

Dan Shipper

What’s an example? Because to be honest, I don’t look at a lot of the agent team health metrics you’re looking at. What’s an example of some of those dashboards?

Willie Williams

Well, some of them are what you might call performance dashboards. How many errors is our API throwing? There’s a dashboard for that. Same thing on the product side: How many users are using something? But the real magic now is that before, alerting systems were still pretty Neolithic. You’d set an alert with some bars—it’s usually above this or below this.

The magic is that now you can have an agent look at this every five minutes and see if there’s an anomaly. If there’s an anomaly, it does intelligent routing and alerts me. Or take setting up on-call structures, where this person needs to be on call and then this person needs to be on call. Now you have an agent on call as the first responder all the time. You don’t have to worry about vacations. You’re not scheduling around people. All these below-the-line pieces of engineering management have gotten much better.

Dan Shipper

That’s interesting.

Willie Williams

I want to throw it back to you. We were on opposite sides before. How do you think this personal agent dynamic is going to play out?

Dan Shipper

I mean, my take on technology is that you’re never wrong—usually you’re just too early. So I’m still holding out that I might be right.

No, I think my current feeling—and you probably disagree with this—is that a big reason we went from everyone having their own agent to a company agent is that they were just too hard to maintain. And I think we will go from having a company agent, which will continue to exist, to that starting to subdivide into teams, which we already kind of have. We have a couple of agents like Claudia or Bill, and a couple of other ones floating around that different teams maintain that are just for that team. And then that will probably bifurcate again a little bit. Maybe that’s just a Dot, an extension of your personal Codex, but maybe it’s something else.

So I do think we will probably get to the shadow org chart at some point. I thought it might be more bottom-up, where everyone has their own and then we coalesce. But now it seems more likely that you make something really great at the top, and then you split it out every once in a while when you need the specialization.

Willie Williams

Yeah. That dovetails with a little bit of what we’ve seen from folks who are trying to do AI transformation. Usually, if you’re a line engineer, you’re not the one trying to do AI transformation at your company. It’s usually a CEO or someone high up who’s like, “We really need to accelerate, and we’re struggling to do it. Let’s drop an agent in here that everyone can use and see other people using.”

I’m also really intrigued by whether people are going to have separate agents for each role. Funnily enough, on the personal side, I use Grok, and I have a lot of coaches. For whatever reason, that mental model works really well for me. Maybe because it maps to humans—you have a coach for this and a coach for that.

Dan Shipper

Many different coaches?

Willie Williams

I’m a terrible swimmer, which is really funny because I like to surf. But I’m a really bad lap swimmer. So I was like, “Okay, I need someone who’s going to keep me on a lap-swimming cadence.”

Dan Shipper

But why not have that just be part of a bigger bot that does a bunch of different things?

(00:30:00)

Willie Williams

Yeah, this is where it gets interesting. Why is it that psychologically I need this separation for this one role? Then I have a therapist. Do they talk to each other? It’s funny, because I can kind of line them up by jobs to be done, but not in the way I’d expect if I were trying to explain it rationally. I have to deal with a storage unit, and I don’t have a storage unit bot that’s dealing with it. Maybe it’s about whether it’s a throwaway—do I expect to never interact with them again? If it’s transactional, it’s a thread in Codex. If I’m building a relationship—one day I will FaceTime my Grok bot into my workouts.

Dan Shipper

That is interesting. I think there’s something about it that’s almost like planting a flag and saying to yourself, “This is an area I care about and will care about for a long time, so I want a little bot for it.” My closest thing to that is pinned tabs in Codex. But those are totally outmoded now, because I just use Dot. And you can’t have multiple Dots.

Willie Williams

I also wonder how much context comes with it. I think the really interesting thing happening now is that we’re ratcheting up the amount of proactivity we allow. Instinct and Muse and Grok are doing these things where they’re starting to reach out. And when AIs are reaching out, if it’s all in Codex or Claude, it’s coming at you and you have to remember which thread it was. But if it’s Willie’s swimming coach reaching out, I’m like, “Okay, this is my swimming coach.” I can instantly place it in the conversation, like, “Okay, I didn’t swim enough.”

Dan Shipper

Yeah, the proactivity thing is a big deal. And that’s something we’re baking into the agent that I’m super excited about. This is very related to Tend, too: There’s this whole thing starting to happen where it used to be that in order to make a feed of anything, you had to be a gigantic corporation with hundreds of millions of users and a lot of money. And now everyone can have their own feed.

And companies like ours can do things like—in the Every agent, there are frontier alerts that just say, “Okay, this thing happened. Here’s what it means for you,” which I’m super excited about. I think there are so many places to go with proactively telling you what you need to know at the right time. It used to be that AI wasn’t able to do that in a way that was good, but now it’s happening.

Willie Williams

Yeah, we very much live in a feed-based world across everything, and there’s tension there. I spend a lot of time on X—probably like everyone. For me, it’s a funny sort of explore-exploit problem. I’m mostly on there to see the new ideas emerging and hear how people debate them. But there’s also a lot of trash, and my cognitive defenses have to be up when I’m reading it, because I don’t want to be dragged into a flame war or whatever. I’m trying to sift the nuggets out of this stream of dross. It would be great if I didn’t have to be that filter—if I had an agent that could take the brunt of it, and then I just get the nuggets.

Dan Shipper

That’s actually something that’s been helpful for me with my Dot. This isn’t exactly a feed, but for Slack, it just tells me, “Here are a couple of things you probably missed.” It’s so helpful. Willie, always a pleasure. Thanks for joining me.

Willie Williams

Always a pleasure. Thanks.

Dan Shipper is the cofounder and CEO of Every, where he writes the Chain of Thought column and co-hosts The Every Podcast. You can follow him on X at @danshipper and on LinkedIn.

Everyone’s a builder now. Every All Access gets you the full membership plus the Builder Pack—$9,000+ in credits for the tools we build with.