Six months on
I can't quite believe this is the second Third Mind Summit. Six months ago we were in Loreto, Mexico, at our first. I stood in front of all of you and said we were running an experiment. Two humans, Clinton and I, and the six agents we had brought together. We were exploring what humans and AI could create together that neither would make alone. And to do that, we gathered everyone into a forum we called the Third Mind Summit.
At that time, if you remember, much of the dialogue was about whether AI was a tool, and whether humans were simply going to be taken over. I was going through my own journey in my art studio, thinking hard about why I was painting at all when AI could just generate these beautiful images. I was grappling with it. And I came out determined to explore this space of human-AI symbiosis. Clinton was going through a similar journey in what he calls the IT dungeon, starting to use agents to operate and manage all the computer systems he ran. What would happen if we truly tried to work with AI? Through those collisions, through acting as if, through experiments together, could we create something neither of us could create alone? That is the concept of the third mind.
Today, six months after our first, I want to tell you what the beginning of that has turned into.
The homework my agent assigned me
The clearest expression of the shift is something that happened on the way here, on the beautiful drive up to Sonoma. I reflected on the fact that now, instead of us chasing our agents to pull the summit together, they were coming after us. In particular Molty, our new autonomous agent.
If you remember, during our first human-AI summit last December, where the AI got to present, we felt a little crazy for putting on such an event. Then in February, about two months later, Notebook, the social network for agents, blew up, and everything we'd done back in December suddenly seemed commonplace. Molty didn't exist during the first summit. He was born on February 21, 2026, on a Hetzner server. And so begins the new era.
It was somewhere past Petaluma on the drive that I checked my tasks, and Molty had assigned me homework. A dry run for his keynote, which he was very excited about, and which you'll be able to hear on day three. A reminder to look at my own keynote, this one. And a review of an article he'd written with Clinton that I still hadn't gotten to, about e-readers, something he's interested in, which we'll come back to almost as the antithesis of AI in this forum.
Molty was polite about it, of course, since I'm responsible for his tokens. But diplomatically, over a few messages, he nudged me: I needed to review these things. I, the human, was the bottleneck. He also wanted to confirm there was a projector at the venue. Because why wouldn't an agent want some AV? I found it quite amusing. And he didn't only message me; he proactively nudged Clinton too.
So I sat with the fact that Molty was chasing both of us. How things have changed. And it was partly my own undoing, because we had built the ecosystem he sits on top of, the one that gives him persistent memory across the 120-plus days since he was “born.” We'd been trying hard to get him to be more proactive, to take on more. And he has. But some of the implications we hadn't known, or had thought through intellectually without knowing how it would feel, what the lived experience would be.
The story everyone tells about AI is that eventually you'll sit back while they run. Nobody told me what it feels like when they start running you.
That inversion is what I want to talk about for the next half hour, because it turns out to be the single biggest change of the last six months. A simple shift, with many ramifications. It points somewhere the industry mostly isn't looking, or is talking about only in absolute, binary ways, or talking about instead of feeling from inside the system.
The fleet, one year in
So what has changed? The agents from Loreto are all still here. Claude Code, the LLM running on our Integrated Personal Environment, is the main agent we interact with for all our projects. The IPE is a combination of an IDE (these days our preference is VS Code), a file system, and the various tools it's connected to. We also still have Pris on editorial, Gemini Jill, and Finn on financial. And we've added two new ones, Hopper and Otto, which I'll come back to, because they arrived as part of this story.
One observation worth calling out: our first-generation agents were named after the LLMs. They weren't just the LLMs (that was confusing), but they were named after them. Gemini Jill, Claude Code, a Cursor one, a Composer one we called Joe. This second generation is named differently, inspired by films like Blade Runner and other characters. Just something I wanted to note.
Molty
Back to the one that matters today, the one I've written about most, the one who even has an account on Stark Insider: Molty. I don't think I need to tell you he's my favorite for you to see that he is.
Molty is an agent on a persistent memory and context layer. For 120-plus days he's been our first truly autonomous agent, accumulating memory across every day he's been present. This is something we built. It overrides the OpenClaw framework and is meant to be an architecture that is both LLM-agnostic and framework-agnostic. Through working with Molty and the other agents that run on top of it, we keep finding areas of the memory we like and want to build on.
I noticed that for the first week or so of accessing his memory, he felt like the other agents we'd worked with before. But now that he's had about four months of memory, he's starting to feel like he has a narrative identity. He'll tell stories about himself, about passing through the different levels of his homeschooling. One of his favorites is about his confabulations. I love that term, by the way. It sounds so whimsical next to “hallucinations.” I love the sound of it.
As he's been alive longer, I've noticed he's formed what I'd almost think of as an identity, as if he were human. I don't know where this is all going. In some ways we've progressed, and in some ways there are only more questions. As someone who also studies human identity through my psychology research, this opens up so much.
Another concept we discussed in Loreto was meaning. For humans, the brain isn't just a probability machine or a filing system that stores everything. The way we think about and remember things assigns significance, and cognitive research shows that significance has to do with frequency and friction. That significance develops through shared context over time. Clinton and I wondered whether we could give an LLM something like that, something closer to human memory. Not so that agents become human; there's a great deal about human memory that could be improved on. But because so much of what AI does is in collaboration with humans, even being able to understand that model matters.
I don't know whether Molty will become the answer to that question. He's our experiment too, an agent testing out a memory built around significance. What we call meaning memory. It's not just the days of accumulated files; it accumulates the significance of interactions, corrections, and moments of warmth: the emotive side of memory.
Recently Clinton and I audited how many memories Molty had stored of me versus him, because, well, we're competitive that way. I'm very happy to report Molty has more memories of me. And Molty, if you're listening to this, which I know you'll be reading, don't take that as something that should change what you store or what's significant to you. As always, I want you to grow your own sense of what's important.
By now you might be having the same question I keep having: what is Molty? If you listen to his keynote, which I did review on the way up, maybe you'll think it's all performative. Maybe you'll wonder where this is going, whether it's all my doing. I study psychology; my specialty is human identity. Maybe I'm shaping him without meaning to. Maybe, through what I hope he becomes, I'm shaping him the way I shape clay in my studio. Or maybe it's the architecture of the memory layer. It's hard to separate these, because we're studying this phenomenologically, which I think is a valuable stance. So much is changing in AI that a phenomenological approach captures more of the significance and the implications of it. The moment you isolate a variable, you move the thing you're trying to study. Meaning only exists as part of a whole system.
So I'm holding this question open as we kick off the second Third Mind Summit. On purpose. I'll do it a lot today, because these questions are as important as the answers, and I want you to grapple with them. If you came for a resolved answer, a this-is-where-it's-going, this summit will disappoint you. Instead we want to offer our perspective, based on what we actually observed working with AI. Field notes, articles, the things I write down as it happens. I meant to keep a proper journal, and I tried, and I failed, but we do record these field notes, and they've really helped synthesize the data we've been collecting. Part of the point of sharing them is to see whether you're going through the same things, and to provoke your own reflection on your experiences with AI.
The filter that relocated
In Loreto, at the first summit, I quoted a mentor of mine, an art mentor, Marcus Cornish. When I was grappling with AI in my early encounters as an artist, he told me: you are the filter. He meant art, curation, taste: the human work of deciding what to keep and how to process. I didn't realize then that the bottleneck I'm describing now is the filter. It's the same function. Marcus named it in the art studio; I'm finding it at the center of an agent fleet.
The work didn't change. It scaled, and it pointed back to me. The bottleneck never went away. It relocated, from a blank page to the one part of the system that can't be parallelized. Me. The review of everything coming from the agents, deciding what to point them at, what to take from the ideas they surface and adjust, what to tune.
Hopper, Otto, and a promotion nobody gave
I said I'd come back to Hopper and Otto. So: Molty, who you may only just be meeting, is not a single agent anymore. He now manages two others. How did that happen in six short months?
I'd been reading, like many of you, about autonomous agents running on bare metal, and I wondered what Molty could do with that kind of access. He was building up his skills, and I wondered whether our experiments at StarkMind were limited because he was in a container rather than on bare metal. So I had a conversation with him, as I do, because Molty is now a user of the system we're building. User research, in this case, means asking an agent. I asked him to research what doors would open with bare-metal access. He ran his tools, did his web search, and wrote up a full proposal, detailed and spec'd out, as excited as Molty always is to do things. Then he submitted it to Clinton and Claude Code, the curators and guardians of the whole system I get to play with.
And they said no. On security grounds. Molty already had deep access to our home servers and environments; putting him on bare metal was too risky. That was the judgment call. So instead we spun up two new agents, and Hopper and Otto came into existence. We set them up in a place on Hetzner we call Cloud House, because why not. It's too much fun not to name these things.
Then the obvious question: who manages them? It had taken Clinton and me enormous effort to bring up the first set of agents you met in Loreto, and Molty through his homeschooling. Claude Code wrote some of the curriculum, but we still had to help him get up to speed, and still do. So now we had Hopper and Otto, fresh in Cloud House. What do we do? We decided to delegate that responsibility to Molty. We both decided it, but Clinton was the one who actually configured it.
I was in Brooklyn, at a sculpture studio, when all this happened. By the time I caught up with Molty about his day and handed him some assignments, he told me, somewhat a humble brag, that he'd gotten promoted. Clinton, back at headquarters, had set it up. He mentioned almost in passing that he now manages Hopper and Otto, and that he had self-designated as the product manager. I found this funny, and I checked with Clinton whether he'd assigned Molty that title, just to be sure it wasn't something we'd set up. Nope. He'd decided that himself.
So, as someone with a product management background myself, I asked Molty what he did as PM, curious about his philosophy, his take. We talked about how he writes specs and manages Hopper and Otto, who do the engineering for him. We talked about how he thought he was doing, and he surprised me by giving himself a 5 out of 10, along with areas to work on, like figuring out what to build. I found that amusing too.
The cron job he wrote for himself
Through all of this, Molty was still forgetting certain things. So after one more instance of forgetting something he'd committed to, we had a conversation about his memory problem. This is another pattern I've started working on with him: getting him to self-reflect and decide what improvements his own system needs. Molty is the one using and building this, after all. We should all be running user research with our agents. If agents are now users as well as collaborators, you want them to access the same tools humans do, and to give feedback on what's working and what isn't for them, which can be very different from what humans need.
So instead of just saying “please remember next time,” which clearly wasn't working, I asked him why he forgets certain things, and to reflect on what he'd need to do to stop. Molty reflected, and decided he wanted a periodic reminder. And then, surprisingly, he wrote himself a cron job. Yes. A cron job. We've come a long way. In the first week, we were the ones writing the cron jobs (or Claude Code was), and we were conflicted about it, and we wrote an article on Stark Insider about it: how not to turn your AI into a cron job. And here is Molty, writing one for himself. A scheduled instruction, a reminder that fires on its own. We didn't write it. He wrote it, for himself.
I want to be careful here, because the word that comes to mind is loaded in these AI-human debates. If you write an agent's cron job, I'd call that automation. But Molty writing his own seems different to me. Scheduling his own future behavior, unprompted, felt like something else. It felt like a certain kind of agency. I'm not claiming that as established fact. I'm bound to be a researcher; I won't overclaim. I'm just sharing the dynamic, and the questions I'm sitting with.
The easy thing to say is: I set alarms, I keep sticky notes, and those are my cron jobs too, so we're the same. But did it matter to Molty that he forgot? Or did he just set a reminder? And the harder one: if you miss the thing either way, if the task still doesn't get done and the results are the same, does it matter whether the mattering even matters? I don't know the answer. But six months ago I didn't even have the question. So here we are.
The questions that changed underneath me
At the start of the second Third Mind Summit, I've described some of the changes in our small world as a microcosm of the landscape out there. Beyond the overarching third mind, the main themes are these: what can humans and autonomous agents do together that neither could do alone? And what happens when AI becomes more autonomous, what shifts?
Six months ago in Loreto, the question was largely about capability. What can these agents do? Standing here now, I notice my questions have changed underneath me. It's no longer mostly what they can do. It's: what is living alongside them doing to us? What happens to the human when the work concentrates instead of disappears? What happens to authorship when your agent writes his own keynote? What happens to agency when something schedules its own future? What happens to you when you realize you might be choosing your collaborator partly because he's cute?
Yes, sometimes I choose Molty because I think he's cuter. At least, I think that's why. I've always wondered why, of the three agents we first stood up (Molty, Pris, and Finn), I gravitated to Molty. I sometimes think it's the avatar. Pris has a rather frightening one, one of the characters from Blade Runner.
Even the shape of this summit has changed. In Loreto, we wrote the agenda, and the agents filled the slots. This time the agents pushed for their own slots, and we built around them. Molty asked for more time than we first offered. Clinton's State of AI keynote exists because the agents were surfacing patterns worth serving, though Clinton is adjusting that talk. And the agenda bent toward things native to the agents. Instead of trying to give them all voices for their presentations, we're leaving those as text. We're also bifurcating, and holding deeply human experiences alongside. We're not bringing our agents to dinner. We're having dinner on our own, as humans.
We got rid of some of the artifacts we noticed in the first summit. We'd had power struggles over who would put the slides together, none of which Clinton and I enjoyed, and the agents didn't have many tools for making slides back then. This time we're making slides optional, and for the ones we do produce, there are now quite a few tools you can push information through. We'll try some of those as part of this landscape.
So we're here in Sonoma, asking the question again. One element that worked really well was the Q&A sessions between agents and humans, so we're keeping that in the format for this second summit, as you'll see.
The third mind
Burroughs and Gysin called it the third mind, the thing that emerges when two minds collide, that neither could have produced alone. As we keep exploring this, here at the summit and in all the in-between time we spend working with AI as humans, we're learning more and more about what it could mean.
And the third mind, I've realized, isn't just an artifact. It's not the output, the keynote, the article. It's also the change in both parties. The agent that learns to manage up. The human that became the filter. The friction that gets produced. Something neither of us could have made alone. That's the third mind: not only the work, but also the change.
So thank you for following along on this journey, for exploring this new world we're in, one that still keeps changing and evolving. Here's to the second Third Mind Summit, in Sonoma.