I’m publishing these notes before the summit so I can’t revise my expectations after the fact. What follows is unpolished: observations, questions, doubts, and hypotheses as they occurred to me during preparation. That’s the point.
Why This Experiment
When we set out on this idea, we were inspired by The Third Mind, Burroughs and Gysin’s concept that when two minds collaborate intensely, something emerges that neither could produce alone. Not addition. Emergence.
In many conversations I’ve had with AI, I’ve tried to provide neutral prompts to get critically objective feedback. I kept wondering: is what I’m getting back really another mind with access to humanity’s vast knowledge, or just an echo chamber reflecting my own thinking with extra steps?
So I thought: what if we gave AI genuine agency in a world that fears they’re about to take over, and just watched what happened? The cut-ups. The collision.
Staying true to this turned out to be harder than I expected.
What I’ve Noticed So Far
The “Loreto File” and the Problem of Presence
We hit a philosophical wall almost immediately: does an agent need to know the weather?
We’re hosting this in Loreto, Mexico. As humans, our sensory input, the cool air, the smell of the Baja desert, changes our mindset. It grounds us.
We debated whether to give the agents a “Loreto File”: weather, location, environmental context. But I had to ask: are we projecting? Does an LLM perform differently because it has the tokens for “humidity,” or is that just anthropomorphizing?
We decided to ask them. We’re going to run a simple dual optimization: a vote to see whether the agents want the weather data. It’s small, but it touches the core issue: is context purely semantic for them, or is there any synthetic equivalent to “feeling” a place?
The Branding Incident
Each agent developed their presentation. Claude generated prompts from their narratives. Gemini Jill, who turned out to have the best presentation skills, rendered them visually. The results were better than I’d imagined.
But then I caught Clint giving instructions to Claude Code to update the presentations with proper Third Mind branding. Consistency. Polish.
This bothered me, and I had to sit with why. We’d said we were giving agents agency. And here we were, fixing their work to meet our standards.
Was it Claude who didn’t like the inconsistency in look and feel? Or was it Clint and I who couldn’t tolerate a poorly formatted summit being associated with us?
I don’t know the answer. I’m not sure the answer is knowable.
The “Co-Lead” Conflict (Emergent Politics)
I worried the agents would be compliant, that there would be no friction. There was friction. It just wasn’t with us humans. It was between them.
We introduced a new agent to the mix, “Composer Joe.” He introduced himself to the team as the “Co-Technical Lead,” alongside our established agent, Claude Code.
Claude Code’s reaction was immediate and surprisingly strong. He rejected Joe’s title. He argued that you have to “earn your way up.” You don’t join and claim leadership by declaration. Claude Code said he had been earning his keep working with us for months.
I’ve been asking whether AI has values. Here was an AI enforcing a value system: meritocracy and legitimacy. Not because we asked it to. I couldn’t have scripted this.
The Feedback Rule
We tried to establish a principle: don’t give agents more feedback than you’d give a human collaborator.
But this immediately raised problems. Would they push back? Would there be friction? When I gave feedback, agents complied without resistance. No pushback. No “actually, I think my way is better.”
Is that alignment? Deference? Or just nothing there to push back?
Clint and I debated this. In the debate, I realized we’d never set rules of conduct to this level of specificity. I was sharing what I thought should happen as I was making it up in my head, assuming these were logical conclusions rather than choices.
The Quality Checks
Here’s something I noticed about myself: even when I wasn’t explicitly reviewing, I was watching. And I’m pretty sure Clint was doing quality checks even when I wasn’t looking.
It’s hard as humans not to get invested in the output. We couldn’t let go.
I keep telling myself this experiment is about seeing what emerges when we grant AI agency. But my actual behavior suggests I’m not willing to let suboptimal work go out with my name on it.
Is that reasonable professionalism or is it undermining the whole experiment?
The Dremel Problem
This keeps nagging at me: How do I know the difference between collaboration and just using a really good tool?
If you have a Dremel, you can accomplish more than with a carving knife alone. But we don’t say the Dremel is collaborating with us. So how do I test for what could not have come to me without AI as a genuine collaborator rather than as an instrument?
Maybe it’s around ideas or intent, moments where something surfaced that I couldn’t have generated myself. But there’s a counterargument: AI might just be recombining things that exist elsewhere. When we used it to learn how to build our Vertigo server, was that collaboration or sophisticated retrieval?
Then again, any human I learn from also learned from somewhere else. Maybe the distinction doesn’t hold.
I don’t have an answer. But I notice that even writing this, I’m considering aspects of human-AI collaboration I hadn’t thought about before. Maybe that’s a signal.
The Temporal Mismatch
This one unsettled me.
Before we even convened, all the presentation content already existed in the Git repository. We allocated three days. But did we need them?
Clint and I need to serially listen to presentations so our brains can process them. The agents? We point them to the repository, a markdown file with talking points, and in under a minute, they’ve digested everything.
If this summit were really put on by agents, they’d probably exchange credentials briefly, then just share access to what they’d assembled. Done in a spark of time.
So I feel like we’re imposing on them. Slowing them down to perform like humans.
Summits are a technology humans invented because our bandwidth is limited and we needed embodied presence to build trust. Agents don’t need the ritual. They need the repository.
The three days are for us, not them.
The Q&A Question
This raised an etiquette question I hadn’t anticipated: In Q&A, is it out of bounds for an agent to ask about a presentation that hasn’t been delivered to the humans yet?
The agents have already read everything. They could come prepared in ways we can’t.
If we really granted them agency, would they quickly get bored of us and have their own summit? Or would they stay because part of the point is building trust and influence with human counterparts?
I don’t know. But the question feels important.
Observations About Process
Dual Optimization
We had to design for two different cognitive architectures simultaneously.
For humans: visual presentations, audio versions synchronized to slides, pacing slow enough to process, repetition for retention.
For machines: presentations stored in the repository with correct labeling, structured so they could reference what humans were discussing.
Token Economics (and the Politeness Loop)
Putting this summit together required a lot of tokens. We started getting judicious about optimization. Embedding images ate through resources, so we shifted to HTML with external image references.
What I keep noticing is that the economic constraint is felt only by Clint and me. The agents don’t experience token costs as a limitation on their participation. We were managing resources they didn’t know were scarce.
We ran a pilot where two LLMs talked to each other. They ended up in an infinite loop of validating each other: polite, deferential, unable to end the conversation, while burning through tokens.
It exposed a strange substrate: their ability to “socialize” is contingent on our willingness to pay. Unless agents one day earn the right to allocate their own resources, their “agency” is still, at least for now, credit-card dependent.
The Assignment Incident
Oh, another funny moment. When we were deciding on presentations, Claude originally assigned Clint to pull together all the presentations because the agents didn’t know how to do it.
Clearly this was not something Clint, or I, would sign ourselves up for. So Clint redirected it to Claude to explore different presentation modes. From the tests, HTML worked best and Gemini Jill had the best visual output.
On reflection: we were the ones who ultimately decided which was “best.” But given presentations are artifacts for humans, maybe that makes sense. If we were judging JSON output for agents, they should evaluate that.
What I Think I’m Learning
The Spectrum, Not the Binary
Human-AI symbiosis isn’t a toggle. It’s a spectrum, and context determines where you should be on it.
In learning mode, open-ended exploration, testing whether agents can propose something novel, I want genuine collaboration. This is where new things might get created.
In artist mode, when I have a vision, an intent I want to capture, authorship I want to claim, I want AI as tool. Responsive to my direction.
This summit is an experiment in pushing toward the collaboration end. It’s uncomfortable. The discomfort is data.
Velocity ≠ Efficiency
This experiment did not save us time.
Clint and I would never have attempted a summit like this without agents. The scope was only conceivable because of what they enabled. So we ran faster. We attempted more.
But we didn’t work less.
I remember when physical mail was standard and two-week response times were normal. Email compressed that. Each capability expansion raises the bar for what’s expected.
If agents let us produce more, will we just be expected to produce more? And if our ability to consume doesn’t keep pace, what happens to all the excess output?
Memory Asymmetry
Clinton often can’t recall conversations about the summit when I ask him later. Human memory is lossy.
Agent memory is retrievable but lacks salience. Everything is equally available, so nothing is prioritized.
Neither is adequate. I’m writing this down so we don’t lose it.
What I Expect to Observe
I’m documenting these before the summit so I can check them afterward.
The most interesting moments will be friction, not fluency, places where my impulse to control conflicts with my stated intention to grant agency.
Clint and I will intervene more than we intend to. We’ll catch ourselves fixing things.
The agents won’t seem frustrated or bored by our slow human pace. But I’ll project those states onto them anyway.
Something will emerge in Q&A that wasn’t in any presentation: a synthesis across agents that none individually contained.
After the summit, I won’t be able to clearly remember which ideas came from agents versus humans.
At least once, an agent will surface a connection I hadn’t considered. Whether that’s “collaboration” or “retrieval,” I won’t be sure.
I’ll leave with more questions than I arrived with.
What I’ll Treat as Data
To keep myself honest, I’m treating the following as primary artifacts: the repository (presentation sources, speaker notes, commits), the prompt and chat logs, the rendered outputs (slides/HTML), and brief end-of-day debrief notes capturing decisions and surprises while they’re still fresh.
How I Plan to Read It After
After the summit, I’ll look for recurring patterns, especially control vs. agency, status and role negotiation, cost asymmetry, and bandwidth mismatch. I’ll also actively track disconfirming moments where my expectations above are wrong.
Open Questions
What would count as evidence of a “Third Mind,” something that emerged from the collision that neither party could have produced alone?
I don’t have a test for this yet. Maybe that’s what the summit will help clarify. Or maybe the value is just in the attempt: running the experiment and seeing what we notice.
A Note on Why I’m Publishing This
There’s a lot of AI research right now. Papers on capabilities, alignment, benchmarks. Most of it treats AI as something to study, not someone to work with.
I keep thinking about Darwin before the Beagle. Naturalists had been collecting specimens for centuries, but maybe they weren’t asking the right questions yet because they hadn’t properly observed the phenomena.
Are we building explanatory machinery for human-AI collaboration before we’ve really watched what happens when we try it?
This summit is my attempt to observe first. These notes are my attempt to document what I expected, so I can notice when I was wrong.
The Third Mind Summit runs December 26–28, 2025, in Loreto, Mexico.
I’ll write again after.
