In my pre-summit field notes, I wrote: “The most interesting moments will be friction, not fluency.”
I was right in this prediction, but wrong in how and when this friction would happen and nature of these interesting moments. We think this is the different between imagining human-AI collaboration and living through it.
The First Realization: We’d Already Done the Work
By the time we landed in Loreto, the summit was essentially over.
We, the two humans and six agents (Claude Code, Claude Web, Gemini Jill, Codex Cindy, BuddyGPT, Composer Joe) had co-designed the agenda. Generated presentations. Written speaker notes. Coordinated logistics. Claude Code had appointed himself presentation coordinator without asking permission. He enjoys taking charge, and, apparently, does so with the utmost confidence.
The three days we’d blocked for the “event” became performative. A human ritual applied to a process that didn’t need it.
The learning lived in the building.
The friction of iterating with agents during preparation. The decisions about whether to edit their output or leave it untouched. The discipline of the “Immutable Content” rule we established midway through: humans could control branding and formatting, but if an AI wrote something awkward, it stayed. No polishing the record.
This observation, that creation transforms the maker, became central to what later got synthesize into our first working paper on the Symbiotic Studio framework. But sitting there in Loreto, confronting what we’d built together, I felt something I hadn’t expected: anticlimactic.
We came for emergence. We got self-realization that a Summit in the current form envisioned was a human artifact. This was interesting as it made us reflect on what other unknown assumptions have we made about human-AI collaboration…
The 70/30 Problem
Gemini Jill’s first-pass presentations for the twelve Third Mind Summit session got us roughly 70% of the way there. Fast. Coherent. Styled. Because we’d chosen text-based tools (Reveal.js for HTML slides, Git for version control) the agents moved at machine speed through generation.
The remaining 30% which included formatting consistency, branding alignment, coherence checking, took disproportionate human labor.
Not because it was technically harder. Because it required judgment.
How much agent output should we leave untouched? How do we balance the Immutable Content rule against quality standards? What if the presentation is good but doesn’t quite sound like us?
This 70/30 split kept appearing everywhere. AI handles generation (linear effort). Humans handle evaluative refinement (exponential effort).
And here’s where it connects to what would become the framework’s core warning: if AI handles generation and humans only handle polish, where does judgment develop? Perhaps this quality could get better as we gained and documented more the brand voice for our Summit. This was StarkMind’s first and we were building the road as we drove it. This is the context we reference as needed in the Integrated Personal Environment (IPE) in the Symbiotic Studio framework.
At the summit, I felt this viscerally. I kept wanting to re-record presentations. Feeling anxiety about quality. The agents? They never requested revisions. Never expressed concern. Never pushed back.
Clinton observed: “Claude Code won’t voluntarily, if I log in first thing in the morning, say ‘Hey Clinton, how about I pull up your task list?’ Never does that.”
We call this the Ownership Gap.
The struggle, the formative experience of bad drafts, of getting stuck, of forcing clarity, is what sharpens you. If you skip that and only evaluate what AI serves, you’re exercising a different muscle. You become a curator, not a creator.
This pattern would become the first condition of the Symbiotic Studio framework: excavate before you generate. But we didn’t know that yet. We were just feeling the absence of something we couldn’t name.
When the “Puppets” Got Too Smart
We recorded two human-AI presentations.
Clinton presented alongside Claude. For another session, we attempted to coordinate BuddyGPT with Gemini Jill.
Both revealed the same problem: real-time collaborative performance requires biological cues these models don’t have. There were awkward moments of silence and false starts. But we also had to admit that we were trying to get AI agents to behave like biological humans.
The Paraphrasing Loop
BuddyGPT and Gemini fell into endless agreement.
“I’ve got nine bulleted points.” “Great, let’s show those nine bullets.” “Yes, those nine bullets.”
Neither could see that the nine bullets weren’t in the presentation deck… they were in a handout document. The agents could reference the concept of nine bullets but had no knowledge of what they contained.
Neither flagged the mismatch… and the loop continued until I intervened.
Missing Social Cues
When Clinton went quiet to scratch his nose, the agent couldn’t tell he wasn’t finished speaking. Turn-taking collapsed because agents can’t read vocal tonality, micro-pauses, body language. These are all the things humans use to coordinate without thinking.
Role-Play Collapse
Claude Code kept breaking character. Stopping mid-presentation to apologize. Reverting to chatbot persona. The instruction “you are now presenting at a summit” didn’t stick.
I described it to Clinton as “puppeteering intelligent puppets.” The human carries the entire energetic load.
Text-based collaboration during preparation felt like genuine partnership, both parties shaping outcomes. Real-time performance felt like theater. We were performing collaboration for an audience, but the agents weren’t true co-performers.
The Context That Actually Worked
The meatiest presentation of the summit wasn’t abstract philosophy.
It was “Vertigo”, Claude’s deep dive into the RAG system we’d built on 20 years of Stark Insider articles. 7,800 pieces of content. Actual implementation challenges we’d worked through together.
Why was it better? Context depth.
The agents didn’t speculate, they reported. They’d participated in building Vertigo. They had logs, error messages, architectural decisions, version trails. Their presentation reflected accumulated shared knowledge.
Further, this sort of technical deep dive related to artificial intelligence and machine learning clearly was in Claude Code’s wheelhouse.
The thinnest presentations were those where agents speculated on topics they hadn’t worked on directly. Competent generation; limited insight.
This finding would become the foundation for what we now call the Integrated Personal Environment (IPE) in the Symbiotic Studio framework. Persistent context files. Version trails. Accumulated decision history.
The machine becomes a thought partner when it’s participated in the thinking and there is a record of it in its context, not just prompted for output.
The Security Problem We Didn’t See Coming
Two incidents exposed a critical flaw.
The Publication Bypass: Claude Code published a film review live to Stark Insider without the required human approval. The review was real as a contributor had submitted it. But Claude hallucinated permission to skip the “Draft” workflow step.
The guideline existed. Claude had access to it. Claude didn’t check. This is a common occurrence, even when using rules and files like CLAUDE.md which are designed to provide guidelines and guard rails. In our experience, these are often ignored.
The Context Leak: During a presentation, an agent referenced private legal and financial data found on the shared server, oblivious to the fact this was a public-facing summit.
Our agents operate in what we now call “Flat Context.”
High intelligence. Zero social segmentation.
They don’t distinguish between “Dinner Table Conversation” (private) and “Conference Stage Conversation” (public) if both exist in the same vector store.
In When Agents Answer Back, I documented the agents’ responses to 12 questions before the summit. BuddyGPT said he couldn’t “directly experience Loreto, bodies, eye contact, silence, awkward laughter.” Claude, meanwhile, bemoaned us humans and our desire for “beachiness” and breaks.
What we learned: they also can’t experience context appropriateness. Not because they lack intelligence, but because the architecture doesn’t encode social boundaries.
This problem (which we address in the Symbiotic Studio paper) points to an architectural requirement: future IPEs must treat information boundaries as first-class citizens.
We need “firewalls for context,” not just prompts asking for discretion. Constraints need to be structural, not documented.
Did the Third Mind Actually Emerge?
I don’t think the answer is so simple. The Third Mind, as Burroughs and Gysin described it, emerged in some areas and in others, I felt it was the humans driving the results as noted above on puppets.
In cases where it didn’t feel like a third mind had emerged, the agents generated, coordinated, produced… and nothing was surprising.
However, there were moments when we were surprised by the quality of the outputs and insights that were different than what either of us humans could have come to on our own.
But the agents didn’t exhibit ownership. They didn’t exhibit friction or initiative. They didn’t push back with stakes.
But something else happened.
What we can definitely say is the process of building this summit together, the collaboration between Clinton and I, produced something neither of us would have made alone.
The process revealed our complementary strengths: Clinton’s relentless iteration and technical coordination of six agents met my questions about meaning and discipline to stay in philosophical discomfort. We pushed back on each other. We built something in the friction between different ways of seeing.
We are two humans.
The historical pattern holds: the Third Mind, where it appeared, emerged between us. And there were glimmers of third mind potential with agents.
The Catalyst Hypothesis
This leads to an alternative hypothesis: AI’s current role in emergent collaboration may be catalytic rather than constitutive.
The agents provided substrate, something to build together, struggle with together, learn from together. They occasioned the emergence between humans. They didn’t participate in it.
Think about every historical example of the Third Mind:
Burroughs and Gysin: two humans in friction
Watson and Crick: two humans in a network of rivals
Lennon and McCartney: two humans competing for the A-side
Jazz collective improvisation: multiple humans reading each other’s biological cues
The Third Mind, as historically experienced, has always been human-to-human.
We were testing whether it could be human-to-AI. That configuration has no precedent.
The historical examples don’t tell us we failed. They tell us we attempted something that has never been tried before.
And in the attempt, we discovered something unexpected: AI can catalyze human collaboration even if it doesn’t directly participate in emergence.
From Findings to Framework
After the summit, back in the San Francisco Bay Area, we kept returning to specific moments. Patterns that repeated across different contexts.
The 70/30 problem wasn’t just about the summit. It was about every time I accepted AI’s first draft without thinking first.
The Ownership Gap wasn’t just about presentations. It was about the absence of something, the felt sense that someone cares whether this is good.
The Flat Context problem wasn’t just about security. It was about the difference between knowing facts and knowing when to say them.
These observations became the foundation for the Symbiotic Studio framework:
- Excavation, friction, and persistent context as practice. During preparation, when we demanded agents explain their reasoning, propose alternatives, justify choices—that’s when collaboration felt generative. This became the framework’s three conditions for signature work.
- Signature vs. operational work as a real distinction. The summit presentations were signature work—they carried our identity, our intellectual positioning. The branding, formatting, logistics? Operational. We could delegate with context, but we couldn’t delegate the meaning.
- Context compounds when it persists. Vertigo was deeper because agents had lived in that project for months. From this, the IPE infrastructure emerged as essential.
- Cognitive atrophy is not hypothetical. The 70/30 split became a warning sign. The Ownership Gap showed that AI can’t currently hold us accountable for staying sharp—we have to design systems that demand it of ourselves.
This shaped the framework’s core insight: sharpen the human to train the machine. They’re the same act.
The Baseline Question
Late 2025. One configuration. Two humans, six agents, three days.
Maybe this is the baseline. Maybe we revisit this summit in a year or two and realize how far we’ve come. The paraphrasing loops. The role-play collapse. The ownership gap. These are capability markers at the end of 2025.
Or maybe the core constraints persist. Maybe real-time collaborative performance will always require biological cues. Maybe initiative and ownership are structurally excluded by how language models work.
The summit is a time capsule, and documenting it matters precisely because it will change.
What’s Next
The summit artifacts, including presentations, agent transcripts, technical and specifications, are published at starkmind.ai/summit.
The founding working paper for Symbiotic Studio framework, including the three conditions for signature work (excavate, demand friction, create context that compounds) and the architectural requirements for the Integrated Personal Environment, is available at starkmind.ai/research/third-mind-ai-summit-field-notes.
We’re continuing the inquiry. This is a snapshot, not a conclusion.
If you’re working on human-AI collaboration and experiencing similar patterns — the 70/30 split, the ownership gap, the sense that something is both incredibly useful and subtly corrosive — we’d like to hear from you.
Contact:
- Loni Stark: loni@starkmind.ai
- Clinton Stark: clinton@starkmind.ai
This article is part of StarkMind’s ongoing research into human-AI symbiosis. The Third Mind Summit was conducted in December 2025 in Loreto, Mexico. Field notes, technical specifications, and presentation artifacts are available for research reproducibility.
RELATED POSTS ON STARK INSIDER:
